Customers and sales: the systems that hold the history of the relationship

This article is also available in Russian: Russian version.

The four loops here answer one question asked from four sides: what do we know about the customer, and who holds that knowledge. CRM holds the history of the deal. The contact centre holds the history of the conversations. The marketing platform holds the history of the touches. The storefront holds the history of the orders. Each works alone. Together they give the thing it was all started for: an understanding of what is happening with the customer.

The problem is shared too, and it is not technical. All four loops are filled in by people, and people fill in what it pays them to fill in. A salesperson does not create the record if the system is a way of watching them. An agent talks correctly instead of helping if speech analytics reads as surveillance. A marketer does not capture a baseline, because without the "before" numbers nobody can argue with the result. A content manager does not complete attributes if nobody owns the product page by name. Hence the identical outcome in all four projects: the system is there, the fields are formally full, the data cannot be trusted.

The second shared thing is the customer identifier. One person easily exists as four records here: a deal in CRM, a number in the telephony platform, a profile in the marketing platform, an account on the storefront. Until those are stitched, personalisation looks like three different offers from one company on the same day.

Below: what each loop does, who needs it, what comes before it. Then how they connect, the vendors you will meet, and a shared checklist.

What's in this article

• Customers and suppliers: CRM and SRM — deal history and the pipeline

• Inbound contact: contact centre, telephony and speech analytics — channels, escalation, conversation review

• Marketing and customer data: MarTech and CDP — the unified profile, campaigns, end-to-end analytics

• Online sales: e-commerce, CMS and POS — storefront, catalogue, order, checkout

• How the four loops connect — the shared profile and the order of moves

• Where AI fits in customer work — what works, where the boundary is, why people resist

• Vendors you will actually meet — common products by subclass

• Common mistakes and a selection checklist — what to check before you buy

Customers and suppliers: CRM and SRM

Sales without CRM look like this: customers live in salespeople's notebooks and phones, a web enquiry waits three days for an answer, and when your best seller leaves, half the base goes with them. If that sounds familiar, this article is for you. CRM is usually the first real system a company owns and where digitalisation of a commercial or service business begins. Let's go through what the class consists of, where to start, why CRM sets your sales process in concrete exactly as it is, and where AI works.

What CRM does — and what it does not do

CRM carries the customer journey end to end: from the first touch — a call, an enquiry, a visit — through the deal to repeat purchases and service. Everything the company knows about the customer sits in one place instead of in people's heads. SRM is the mirror class for the other side: suppliers, procurement, tenders.

What CRM does not do: it does not account for the company's money and does not replace the books — revenue and cost of sales are consolidated in ERP. And it does not sell by itself: it hard-wires whatever sales process exists. With no process you get an expensive digital notebook and sales that do not grow.

Subclasses and modules: what the class is made of

Sales CRM. Pipeline, deals, tasks, touch history. The core of the class.

Effect: deals stop falling through the cracks, the journey becomes visible, and the base belongs to the company rather than the salesperson.

Drawback: it lives on data-entry discipline — empty records make it an expensive address book, and no automation compensates.

Marketing module. Segments, campaigns, trigger sequences against your own base. The boundary with the full MarTech loop is volume: once channels and data multiply, a separate class begins, covered in the MarTech section below.

Effect: repeat sales out of your own base are the cheapest sales you have.

Drawback: on a dirty base it becomes a spam machine that burns loyalty.

Service and support. Tickets, queues, response times, omnichannel — the customer writes wherever is convenient and the company sees it in one window.

Effect: enquiries stop getting lost, and service quality becomes measurable for the first time.

Drawback: speed metrics turn into an end in themselves — you answered fast, but you did not help.

Loyalty. Points, programmes, personal offers.

Effect: keeping a customer is cheaper than acquiring one.

Drawback: the discount needle — a programme built only on discounts trains people to buy only on discount.

CPQ — complex B2B proposals. Product configuration, pricing, quote approval.

Effect: a proposal in hours instead of weeks, without pricing errors.

Drawback: it needs clean product and price data — CPQ on a chaotic price list automates the errors.

SRM — procurement and suppliers. Supplier qualification, tenders, transaction history.

Effect: procurement becomes transparent, and supplier history survives a change of buyer.

Drawback: the class works where procurement is mature — in a small business it is overkill.

Who needs the class, who is too early — and what comes first

Needed by any business where a sale is longer than one touch: B2B, services, complex retail, anything with repeat purchase. Symptoms that it is time: enquiries get lost or wait days; the base lives in spreadsheets and phones; a departing salesperson takes customers along; nobody can say how many deals are in play and at what stage.

Too early or not needed: purely transactional retail with no repeat contact gets by with a till and a loyalty scheme; SRM is overkill while procurement is one person with a phone.

CRM's neighbours are the contact centre and speech analytics (calls become customer history), the storefront and MarTech and CDP: all three feed on the customer base and return events into it.

What it connects to and when to implement. For a commercial or service company CRM is the starting class: its data is born here, at the point of contact. The whole trajectory starts here; documents, reference data, ERP and analytics come after. Forward, CRM hands customer data to the landscape: to ERP for invoices and shipments, to the analytics loop for the sales picture, to MarTech for campaigns. The rule is the same as across the series: process first, system second — CRM accelerates order and cements chaos.

What the class gives the business — and its standard drawbacks

Effect. The fastest of any class in the series: the first results — enquiries not lost, response speed, a transparent pipeline — show up in weeks rather than years. The customer base becomes a company asset. And it is the foundation of the whole customer loop: without CRM data, neither marketing, nor sales analytics, nor AI works.

The drawbacks people talk about less. The class rests on people more than any other: data is entered by salespeople, and if the system means control and overhead to them, the data will be fiction. Configuring it "to fit us" drags on: the pipeline gets rebuilt for months instead of being sold through. And the standard illusion: CRM has been bought, so sales will grow by themselves. They will not — the system shows the process, the manager changes it.

The company CRM is not the executive's personal network

The class has a boundary almost nobody discusses. An executive's personal contacts — people met at conferences, in negotiations, on past projects — do not live in CRM, and rightly so: the corporate base belongs to the company, the personal network to the person. So a senior leader's most under-used asset sits in a phone and a memory: when you need an introduction it starts with "who was it we discussed this with?", and warm relationships go cold without contact.

This is solved by the same mechanism as in sales, only in the personal loop. You need a base of acquaintances with context: who they are, where from, what you talked about, how you could be useful to each other. Plus the discipline of returning to cooling connections. It works even in plain notes, kept honestly; the only thing that does not work is keeping it in your head.

We do this in Seknum, a personal assistant for contacts: it holds the context of each acquaintance and helps you come back in time to connections worth keeping. The honest boundary: Seknum is not a CRM and not a competitor to one. It does not run deals, pipelines or the company's customer base — it works with the executive's personal network, and it is useful regardless of which CRM the company runs, or whether it runs one at all.

2026: the context

CRM is the most crowded and mature class on the map, and the practical question is no longer which system has the features — every mainstream suite closes the basic scenarios. It is which one lands on your sales process and takes root with your salespeople, and where the data physically sits: GDPR made that a board question, not an IT one. The second shift: assistant and scoring features now arrive bundled in the suite rather than as a separate purchase, which moves the decision from "should we buy AI" to "do we let the vendor's model touch our customer data, and on what terms".

Sizing it. Cloud CRM is one of the cheapest entry points in the cycle, but pricing models differ in kind: some vendors charge per user, others per company or per platform, and at a hundred employees the difference is a multiple. Cost is driven by seats, by integrations — telephony, accounting, website, data migration — and by the size of the base you are moving. Cost it after a survey, not off a price page. The most expensive line is not the licence; it is the discipline of filling the system in.

Inbound contact: contact centre, telephony and speech analytics

For a mid-sized business this is often the first "smart" system: AI delivers understandable value here earlier than in production or finance. And it damages reputation fastest here — because it speaks directly to the customer. Let's go through what the loop consists of, who needs it, what comes first, and the boundaries for robots talking to people.

What the loop does — and what it does not do

The loop takes enquiries from every channel — calls, chats, messengers, email — distributes them between agents, stores recordings and turns conversations into data.

What it does not do: it does not replace the customer base. A call by itself is an event; for it to become the history of a relationship you need CRM. And it does not fix the product: if people call because the instructions are unclear and delivery is late, the best contact centre only takes the complaints faster.

Subclasses: what the loop is made of

Telephony and contact distribution. Call handling, queues, routing, recording.

Effect: calls are not lost, load is distributed, conversations are preserved — you have material to review.

Drawback: without integration to the customer base, the agent starts every conversation with "could I take your name, please".

Omnichannel. Website chat, messengers, social channels and email in one agent window.

Effect: the customer writes where it suits them and does not repeat their history in every channel.

Drawback: more channels means more expectations on speed; opening a messenger with no resource to answer it is worse than not opening it.

Speech analytics. Transcription, scoring against a checklist, surfacing problem topics and words.

Effect: conversation quality becomes measurable across the whole stream rather than a sample; topics management did not know about surface.

Drawback: the tool turns into surveillance very easily — and then agents talk correctly instead of helping.

Voice bots and chatbots. Automatic handling of routine enquiries, pre-qualification, simple scenarios.

Effect: routine contacts come off people, and the line holds through peaks.

Drawback: a bot without a fast route to a human irritates more than a queue; here the failure scenario matters more than the main one.

Voice of the customer and feedback. Post-contact scores, surveys, complaints consolidated into themes.

Effect: complaints turn into a list of product problems rather than a stream of irritation.

Drawback: it works only if the themes reach the people who can fix them — otherwise it is collecting complaints for a report.

Who needs the loop, who is too early — and what comes first

Needed once enquiries pass a couple of dozen a day and customers arrive through several channels at once. Symptoms that it is time: calls get lost; at peak, customers cannot get through; nobody knows what people ask about most often; conversation quality is judged by ear from a couple of recordings a month.

Too early or not needed: at ten enquiries a day a phone and an inbox with answering discipline are enough; speech analytics is pointless while volume is low — there is not enough of it for statistics.

What it connects to and when to implement. The loop normally goes in after CRM — otherwise calls do not become customer data and stay a pile of recordings. Input: the customer base and the catalogue. Output: back into CRM (history, tasks), into analytics (topics, load) and into the product (voice of the customer).

What the loop gives the business — and its standard drawbacks

Effect. Availability: the customer gets through and gets an answer — the simplest competitive advantage there is, and one many lose. Measurable service: speed, resolution, contact themes. And product feedback: the contact centre is the only place where the business hears the truth about itself every day.

The drawbacks people talk about less. The loop turns into a metrics conveyor easily: measure only speed and agents learn to close the conversation rather than the problem. Call recordings are personal data, with everything that follows from GDPR on retention, access and the legal basis for recording at all. And robotisation without boundaries damages reputation faster than it saves on headcount.

2026: the context

Speech analytics and voice bots went mainstream within a couple of years: recognition quality rose far enough that the entry threshold dropped to mid-sized business, and the same models now handle several languages well enough for a regional support line. The cost of error rose with it: customers recognise a bot and do not forgive it a dead end with no way out to a human.

Sizing it. Cloud contact centre pricing is driven by agent seats and channels switched on; speech analytics by minutes processed, so the bill scales with call volume rather than headcount. But the real money here is not licences — it is telecoms and agent salaries. They are counted separately and they determine the budget. Cost it after a survey of volumes by channel.

Marketing and customer data: MarTech and CDP

Marketing is the only function that spends money every month and, more often than any other, cannot prove it was well spent. Not because marketers are bad: the customer journey broke into a dozen touches across channels, and connecting the first to the purchase became a technical problem rather than a matter of managerial will. Let's go through what the loop consists of, what a CDP is and why you would want one, what comes before all of it — and why the budget is cut here first.

What the loop does — and what it does not do

The marketing loop collects data on customer behaviour at every point of contact, stitches it into one profile, runs communications against that profile and measures what came of it. Three jobs: collect, reach out, measure.

What it does not do: it does not create demand out of nothing — it amplifies what works and accelerates the failure of what does not. It does not replace CRM: the deal, the relationship history and the salesperson's work live there. And it does not fix the product: the most precise personalisation will not save an offer the customer does not want.

Subclasses: what the loop is made of

Communication automation. Email and messenger campaigns, trigger scenarios, abandoned baskets, reactivation.

Effect: communication stops being a manual campaign and becomes a process that runs without a human in it.

Drawback: the ease of launching scenarios leads to excess: the customer unsubscribes not from one email but from the company. Outbound also has a legal floor — CAN-SPAM, CASL, ePrivacy consent rules — and "we'll sort the consents later" is how a campaign becomes a legal matter.

CDP — customer data platform. Collecting events from every channel and stitching them into one profile: one person, not five anonymous visitors on different devices.

Effect: segmentation based on real behaviour rather than hypotheses, and one shared base for every tool.

Drawback: a CDP does not create data, it unifies it — on three channels with bad tagging it unifies bad data.

End-to-end analytics and attribution. Linking ad spend to enquiries, deals and money; distributing credit between touches.

Effect: it becomes visible which channel brings revenue and which brings traffic.

Drawback: attribution is always a model, never a fact: different models give different answers, and choosing the model is a management decision that for some reason gets taken by agencies.

Advertising management. Launching and optimising campaigns, retargeting, feeding conversion data back into the ad systems.

Effect: budget is redistributed by actual result rather than by clicks.

Drawback: optimising the short distance eats the long one: the system keeps harvesting people who would have bought anyway.

Loyalty and personalisation. Loyalty programmes, personal offers, product recommendations.

Effect: repeat purchases stop being accidental.

Drawback: a discount is the most expensive retention instrument and the easiest one; it papers over problems with a different cause.

Who needs the loop, who is too early — and what comes first

Needed by companies with a mass customer and those who sell through content — particularly once you have three or more contact channels and more customers than a salesperson holds in their head. Symptoms that it is time: the same offer goes to someone who bought that exact thing yesterday; nobody can say where the customer behind your largest deal came from; the ad budget is allocated by feel.

Too early or not needed: in B2B with a dozen large accounts and a long cycle, a CDP is overkill — CRM and a human do the job. Full end-to-end analytics on a single traffic channel is the same: nothing to compute, the contribution is known in advance.

What it connects to and when to implement. The loop normally goes in after CRM and after the digital channels exist: there is nothing to automate marketing on without data about the customer. It belongs to the commercial and service track of the two implementation trajectories; both are in the implementation-order map in the hub article. Input: the website, the contact centre, the storefront and CRM. Output: back into CRM and into the analytics loop. Order in customer identifiers is the territory of MDM: without it, profile stitching gives you three John Smiths instead of one.

What the loop gives the business — and its standard drawbacks

Effect. Controllable spend: the budget stops being a line of faith and becomes a line with a return. Repeat sales: working the existing base is cheaper than buying a new customer, and the loop makes that systematic. And speed of testing: marketing moves from quarterly campaigns to weekly experiments.

The drawbacks people talk about less. The loop creates an illusion of precision: a handsome attribution dashboard looks like a fact when it is a model with assumptions in it. It demands tagging discipline — one agency forgets one tag and a month of data is garbage. And it is the class most exposed to personal data after HR: a customer profile with behavioural history is exactly what a regulator and an attacker both consider valuable.

Why marketing loses its budget first

When a company cuts costs, marketing goes first. The usual explanation is that it is optional. The real reason is different: it cannot show its effect in the terms the CFO counts in.

In Dzhimsher Chelidze's book "Artificial Intelligence. A Practical Guide to Implementation" this is the seventh sin of digitalisation — the absence of a system for measuring effects. The symptoms are named directly: no baseline before implementation, no agreed project metrics, no calculation of return. So is the consequence: you cannot prove value, so the next budget is not granted. The remedy transfers to the marketing loop one to one: measure the baseline before you start, fix the target metrics at the very beginning, report regularly rather than on demand.

The practical translation. Before launching the loop, capture four numbers: cost of acquiring a customer, share of repeat purchases, time from first touch to deal, revenue per customer. Any of them can be disputed, but they give a reference point. Six months later the budget conversation runs around those numbers rather than impressions — and that is the only thing that protects marketing when the cuts come.

2026: the context

Two processes are changing the rules. The first is the tightening of how personal data is handled: consent, retention, liability for breaches. Under GDPR and the ePrivacy consent regime the customer profile became an asset with legal weight, and "where does the data physically sit" stopped being a technical question. The second is the shrinking of third-party identifiers: cross-site tracking is being closed off step by step, and the value of your own first-party data rises. Both push the same way — in favour of whoever collected their base and papered it properly.

Sizing it. Campaign platforms are cheap at small volumes, usually priced on contacts or messages sent. A customer data platform is another order of magnitude, priced on data volume, event throughput and connected sources — exactly what grows once the project succeeds. Basic mechanics take weeks; a loyalty programme months. And do not cost only the platform: content, segments and campaign scenarios are made by people, a separate line. Cost it after a survey of channels and event volumes.

Online sales: e-commerce, CMS and POS

The storefront is the most visible thing a company has, and therefore the most over-rated. The site is launched first, money goes into design and load speed — and then it starts. Half the orders come in against wrong stock figures, product names diverge from the books. And every second return is "what arrived isn't what was in the picture". The problem is not the storefront. Let's go through what the loop consists of, what comes before the site, and why an expensive platform on a small assortment is the classic way to spend a budget for nothing.

What the loop does — and what it does not do

The loop turns interest into a paid order: it shows the catalogue, takes the money, passes the order into the books and to the warehouse, and carries the buyer through to receiving the goods. In physical retail the till does the same: registers the sale, issues the receipt, reports it where the local tax regime requires, posts it to stock records.

What it does not do: it does not create demand — that is the marketing loop and CRM. It does not tidy up your products: the storefront shows exactly the catalogue it was given, duplicates and empty attributes included. And it does not replace logistics: the promise "delivered tomorrow" lives in the warehouse, not on the website.

Subclasses: what the loop is made of

E-commerce platform (CMS). Catalogue, product page, basket, checkout, customer account.

Effect: your own sales channel, where you own both the rules and the data about your buyers.

Drawback: a platform needs traffic — a site with no source of visitors stays an expensive business card, the most common first-year mistake.

Marketplaces and external channels. Selling on someone else's platform, plus tooling to manage assortment, prices and stock there.

Effect: traffic arrives immediately, with no investment in acquisition.

Drawback: you rent the buyer rather than acquire them: commission rises, rules change, and the customer base stays with the platform.

Tills and POS in retail. Registering the sale, the receipt, card acquiring, stock accounting at the point.

Effect: the sale is recorded automatically instead of copied into a notebook; demand data appears the same day.

Drawback: the till is a mandatory requirement rather than an improvement, and it is easy to treat it as a chore while ignoring the data it collects.

Payments and delivery. Acquiring, instalments, carrier integrations, pick-up points, returns.

Effect: payment and delivery stop being manual operations and stop holding the order up.

Drawback: every new payment and delivery method is a separate integration and a separate returns scenario; their number grows faster than revenue. Card payments also drag PCI DSS scope behind them — a reason to keep card data out of your own systems wherever the provider allows.

Product content and stock. Descriptions, attributes, photographs, synchronisation of prices and availability between storefront, warehouse and books.

Effect: the buyer sees what actually exists and understands what they are buying.

Drawback: the most labour-intensive and least-liked work in the loop — and precisely where everything else breaks.

Who needs the loop, who is too early — and what comes first

Needed once sales move online: orders arrive from messengers and the website mixed together, the assortment no longer fits a price list, and there is more than one retail point. Symptoms that it is time: stock on the storefront and in the warehouse diverge; salespeople re-key orders into the accounting system by hand; the same item is named differently on the site, in the books and on the marketplace.

Too early or not needed: if the assortment is a dozen items and sales go through personal contacts, the platform will not pay for itself; start with a presence on an existing marketplace. Building your own storefront at a turnover that will not pay for a support team is not an investment, it is an obligation.

What it connects to and when to implement. The till stands apart: it is a compliance obligation, it goes in immediately and waits in no queue. The storefront goes in after CRM and after a master record for the product exists — with a large assortment that is a separate product content management system, with a small one, discipline in the reference data suffices. The rule is simple: a storefront without a clean catalogue produces returns and chaos. Of the two implementation trajectories this loop belongs to the commercial and service one; both are in the implementation-order map in the hub article. Input: the catalogue and the customer base. Output: orders and demand data into the accounting and analytics loop, picking tasks into WMS. Order in reference data is the territory of MDM and PIM.

What the loop gives the business — and its standard drawbacks

Effect. Sales without limits of time or geography — a platitude that stops being one when you compare revenue before and after. Real-time demand data: what people look at, what they abandon, what they return. And manual labour removed from order intake — the salesperson stops being a transcriber.

The drawbacks people talk about less. The loop exposes everything that was hidden: inaccurate stock, mangled names, delivery promises that cannot be met — a live salesperson used to smooth that over, now the buyer sees it. Cost of ownership grows invisibly, through integrations: payments, delivery, marketplace, accounting — each lives its own life and breaks at its own moment. And dependence on someone else's rules: a change in a marketplace's commission can turn your economics over faster than you can reprice.

Why an expensive platform does not sell

Dzhimsher Chelidze's book "Artificial Intelligence. A Practical Guide to Implementation" sets out seven sins of digitalisation, and the first is the absence of any understanding of what the technology is and where it applies. The system is bought "because it is fashionable" and expected to replace whole departments. The result, in the book's phrase, is "a Ferrari for hauling potatoes". An expensive solution for a task that does not require it.

In online retail this sin is recognisable. A company with two hundred SKUs buys a platform designed for hundreds of thousands, pays for implementation and support — and gets a storefront lacking the one thing it needed: visitors. The remedy is equally direct: work out which task exactly the technology closes, put the leadership through the learning, and pilot with an effect assessment instead of buying off a presentation.

The practical translation for this class: before choosing a platform, answer where the first thousand visitors will come from and who will maintain the product pages. With no answer to either, the platform is not your first move.

2026: the context

The marketplace share of online trade keeps growing, and for many companies the question is no longer "site or platform" but "how do we live in two channels without multiplying our reference data". At the same time regulatory weight on the loop increases — product traceability rules in several categories, card payment security under PCI DSS, consumer protection on returns and price display. The loop stopped being purely commercial and became partly regulatory. I am not putting market-share numbers for e-commerce channels into this text: sources disagree, and putting a number in an article without a verified reference means substituting the plausible for the true.

Sizing it. A templated shop is cheap and quick; a mid-sized one with integrations is another order of magnitude, and cost there is driven by the number of integrations — payments, carriers, marketplaces, accounting — and by catalogue size, not by the licence. The till is not a one-off either: around the terminal come acquiring charged as a percentage of turnover, hardware replacement cycles, and whatever recurring fiscal or tax reporting obligations apply in each country you sell in — and those differ enough between the EU, the US, the UAE and Singapore that a multi-country rollout is costed country by country. If your goods fall under traceability rules, add scanners at each workstation and document exchange with suppliers and the state system. Over a few years the recurring part outweighs the terminal. Cost it after a survey of markets, channels and SKU count.

How the four loops connect

There is more confusion between these classes than anywhere else on the map: functions overlap, and almost every vendor promises to close the neighbouring ones. CRM sends campaigns, the marketing platform tracks deals, the storefront holds a buyer profile, telephony writes records. Separate them not by function but by whose instrument each one is and which decision it serves.

Who manages what. CRM is the salesperson's instrument: it runs a specific deal with a specific person. The contact centre is the service function's: it serves a flow of enquiries at a promised speed. The marketing platform is marketing's: it works with segments, not people by name. The storefront is the instrument of selling without a salesperson: there the buyer decides alone. Hence the different requirements: CRM has to be convenient for the salesperson, the storefront for the buyer, the analytics for the executive.

The order of moves. Almost always you start with CRM: it collects the relationship history everything else stands on. A contact centre with no CRM integration does not create customer data — it connects lines, which is why that link is mandatory rather than optional. The marketing platform goes in once there are more than two or three channels and at least a year of purchase history: before that there is nothing to unify. The storefront is launched after the catalogue is in order — otherwise the site shows buyers exactly the mess sitting in the reference data.

A shared input — the master record. All four loops rest on two reference sets: customers and products. The master customer and the master item live not here but in the data and documents loop. The reference data gives the item itself. Product attributes — dimensions, shelf life, composition — come from the same place: without them you build neither the storefront nor the storage rules. A platform bought before order in the identifiers will unify what unifies and multiply the rest.

A shared output. Downward: orders into the accounting loop and picking tasks to the warehouse. Upward: metrics into analytics — cost of acquisition, share of repeat purchases, deal cycle time, revenue per customer. Those four numbers assemble from all the loops at once, and they are worth capturing before implementation: without a baseline any project result is a matter of faith, and the marketing budget is cut first.

What happens when the order is broken. The company buys a customer data platform without deciding which identifiers count as one person. It launches the storefront before the catalogue. It puts a bot on the first line with the route to a human hidden. It switches on speech analytics by decree. The technology works in every case — but what it collects is not data, it is an imitation of data.

Where AI fits in customer work

This is the loop where AI delivers a fast, well-measurable effect — and the one where it fails most often for reasons that are not technical.

What AI already solves. In sales: transcription and analysis of calls, draft emails, deal scoring, filling records from correspondence — removing part of exactly the data-entry drudgery the class rests on. In the contact centre: review of every conversation rather than a sample, prompting the agent live, bots on routine enquiries, summarising a conversation into the record. In marketing: churn and purchase-propensity prediction, segmentation a human would not have invented, text variants for tests, contact timing. In online sales: generating and normalising product descriptions — the most labour-intensive part of the loop — semantic search instead of word matching, recommendations and search by photograph.

What is coming (a forecast, not a fact). Agents that carry a routine deal or enquiry up to the handover to a human. An assistant for the head of sales — pipeline review and a weekly plan per salesperson. Buyer-side assistants that compare offers and place the order for a person: at that point the product page is written for a machine no less than for a human.

The conditions without which it does not fly. Data: populated records, recorded calls, at least a year of purchase history, a clean catalogue with attributes, correct profile stitching. Process: honest pipeline stages, escalation rules from bot to human, a rule for what to do with a churn prediction, and an answer to who approves a generated description. People and security: everything processed here is personal data, and a model's access to it is granted with the same rigour as an employee's. Where the model influences a decision about a person, the EU AI Act frames that as automated decisioning — which means a named human accountable for the outcome, not a note in the vendor's documentation.

The main cause of failure here is not the technology, it is people. Salespeople and agents receive call analysis as surveillance. What follows is predictable: calls "accidentally" bypass recording, records empty out, agents talk correctly instead of helping. In Dzhimsher Chelidze's book "Artificial Intelligence. A Practical Guide to Implementation" this is trap No. 4 of six — people against the machine. The remedy is from the same place and it is counter-intuitive: reward use of the system, not the result. On the cases in the book, paying for the interaction itself removes the fear of making a mistake, and people stop hiding data. Plus an open conversation about why all calls are being listened to — before wall-to-wall analytics is switched on, not after the first scandal.

Three boundaries worth knowing before the pilot.

First: the model reproduces the bias in your data. Scoring trained on a history where salespeople worked the large accounts and small ones hung unattended will honestly learn that small accounts "don't buy" — and bury them. The same in marketing: the optimiser narrows the audience to people who would have come anyway, while growth requires the opposite. Hence two mandatory procedures: test the scoring against deals it rated low and you nevertheless won, and ring-fence a share of budget for finding new segments, out of the optimiser's reach.

Second: here AI speaks on the company's behalf. A bot that confidently promises a customer something the company does not do creates a reputational and sometimes legal problem, not a technical one. The book lists this among the unacceptable events — a public hallucination in official communication. Hence the mandatory pre-launch check: the failure scenario, that is, what happens when the system is unavailable or has answered nonsense, and who picks up the work.

Third: generated product text is a commitment. A description with a non-existent attribute is not a stylistic error, it is grounds for a return and a complaint. The model does not know your product; it knows how people usually write about products like it. Attributes are taken from the record, not composed, and spot-checking is always required.

Standard failures. Personalisation is switched on over data where one person exists as three records — and the customer gets three different offers from one company on the same day. A bot is put on the first line to save money and the route to an agent is buried at the end of the menu — the saving turns into churn nobody connects to the implementation. Descriptions are generated for the whole catalogue at once — some turn out plausible but wrong, the errors spread across marketplaces, and unpicking them takes longer than writing them by hand.

Vendors you will actually meet

The table lists products you will meet repeatedly on international projects. It is a map, not a ranking: no market shares, no prices, no "leader" claims, because the answer to "which one" depends on your process, your channels and your landscape. Availability, data residency and contractual terms differ by region — check them for your jurisdictions before shortlisting.

• CRM — Salesforce · Microsoft Dynamics 365 · HubSpot · Zoho · Pipedrive · SAP Sales Cloud

• Telephony and contact centre — Genesys · NICE · Five9 · Amazon Connect · Talkdesk · Twilio

• Service desk and customer support — Zendesk · Intercom · SAP Service Cloud

• Marketing automation and campaigns — Adobe Experience Cloud · Braze · Klaviyo · HubSpot

• Customer data platforms and event collection — Segment · Tealium · mParticle · Salesforce Data Cloud

• E-commerce platforms — Shopify · Adobe Commerce · commercetools · BigCommerce · SAP Commerce Cloud

The selection question starts differently in each loop. In CRM it is less about features — every mainstream product closes the basic scenarios — and more about what will take root: an interface salespeople will not sabotage, and integrations with your telephony and website. In the contact centre it starts with channels: where your customers actually are and what you are prepared to serve at the promised speed. In marketing, with the number of contact channels and the quality of your tagging. In online sales, with two numbers: how many SKUs you have and where traffic will come from. Everything else is derivative.

Common mistakes and a selection checklist

The mistakes in the four loops differ in form and are identical in nature: almost all are about people and data, not software.

1. Implementing a system without a process. CRM multiplies what is already there: order stays order, chaos stays chaos. The same for any of the four loops.

2. Compelling instead of motivating. A system used under duress fills with fictional data: formally complete, not to be trusted.

3. Buying off a feature list. The winner is not the most functional system but the one people actually use.

4. Configuring for months. You do not need a perfect pipeline at the start — you need a working one: launch something simple, refine it on the facts.

5. Putting in a bot with no route to a human. A loop without escalation is the most expensive way to save on the first line.

6. Measuring only speed. The conversation closed quickly, the problem stayed: the metric is green and the customer is gone.

7. Opening channels without the resource. A messenger answered a day later is worse than no messenger.

8. Turning conversation analytics into surveillance. Agents will talk correctly instead of helping.

9. Buying a customer data platform before order in the identifiers. It unifies what unifies and multiplies the rest.

10. Not capturing a baseline. Without "before" numbers any result is a matter of faith, and the budget is cut first.

11. Handing the choice of attribution model to an agency. That is a management decision: it determines which channel is recognised as effective.

12. Launching the storefront before the catalogue. The site shows buyers precisely the mess that is in the reference data.

13. Living on a marketplace without your own base. Commissions and rules change, not in your favour, and the buyer stays with the platform.

14. Not designing the return. The scenario for returning goods and money is designed at launch, not at the first complaint.

15. Not returning the voice of the customer to the product. Complaint themes with no addressee are irritation collected for a report.

Checklist before choosing a system

1. The process is written down: sales stages with owners and deadlines, enquiry scenarios with escalation rules, scenarios for returns and partial delivery. Before you choose a system.

2. Four baseline numbers are captured: cost of acquisition, share of repeat purchases, deal cycle time, revenue per customer. Otherwise there is nothing to prove the effect with.

3. The profile-stitching rules are written down: which identifiers count as one person. A condition for all four loops, not the task of one project.

4. An owner is named, by name: of the sales process, of service, of the marketing loop, of product content. Not the contractor and not IT — otherwise nobody fills in the records.

5. The pilot runs in one department or zone with a measurable criterion: share of completed records, response time to an enquiry, share of resolved contacts.

6. Integrations are tested on live data, not in a presentation: telephony and website with CRM, storefront with accounting and the warehouse, contact centre with CRM.

7. The legal side is settled: consents, retention, where customer data physically sits, what leaves for external services, and which of it falls under GDPR, ePrivacy consent rules and PCI DSS in your markets.

8. There is a motivation plan: what people get from the system besides being controlled. I would walk away from a project where this point is empty — it produces completed fields and no data.

9. The export is tested: how you take your customer base and catalogue with you if you change systems.

What's next

The overall map of classes is in Enterprise IT systems: the map. Other pieces in the cycle: Production, assets and warehouse · Money, planning and analytics · Data and documents · Routine automation and AI · People and the IT function · IT infrastructure.

To go deeper, see Dzhimsher Chelidze's books "Digital Transformation for Directors and Owners" and "Artificial Intelligence. Freefall" — free download. If you still have questions, you are welcome to come to us for training or consulting.