Business

What AI Can Actually Do for Your Business, Function by Function

Most AI conversations inside a company start with a technology and then hunt for somewhere to put it. A more useful starting point is the org chart. Every function produces a stream of repetitive judgement, and that stream is where AI either returns more than it costs or quietly does not.

What follows is a map by function rather than by technology: what is genuinely automatable today, what measurable effect to look for, and the conditions under which the investment does not return. Any figures given are the order of magnitude typically seen in practice, not the output of a study. The only number that matters in a specific company is the one measured there before the work starts.

Sales and client work

Three tasks absorb disproportionate commercial time. Proposals can be assembled from what already exists - the CRM record, comparable proposals for similar scope, the current price list, standard terms - so the salesperson edits rather than authors. Inbound enquiries arriving across several channels can be classified against qualification criteria, routed, answered with a first reply, and flagged when they do not fit at all. Meeting preparation becomes one page assembled from the full client history: previous orders, open tickets, payment behaviour, unresolved complaints, what was promised last time.

What to measure is elapsed time, not keystrokes: hours from enquiry to proposal sent, first-response time, the share of enquiries with no human contact after twenty-four hours. Drafting a standard proposal typically moves from hours to tens of minutes. The second-order effect matters more - senior sellers stop spending mornings on requests that were never going to convert.

Where it does not pay: a dozen large bespoke deals a year, each negotiated over months, contain nothing repetitive enough to automate. Nor does any company where client history lives in personal mailboxes rather than the CRM. That is a data discipline problem, and AI on top of it produces confident summaries of incomplete facts.

Finance and accounting

Finance has the clearest arithmetic, because the manual effort is already counted in hours. Document matching - invoices against purchase orders, delivery notes and bank lines - can run on amount, counterparty, date and contract, escalating only exceptions. Incoming documents arriving as PDFs in dozens of layouts can be extracted into structured fields, which removes both the keying and the typing errors that surface weeks later as payment disputes. Receivables follow patterns: a model that has seen a counterparty's history flags invoices that are drifting before the debt ages into a collection problem.

In management reporting, assembling numbers is an ERP and BI task. What AI adds is the commentary - why margin moved, which customers or lines drove it - drafted from the same data and reviewed by the finance lead. Measure the share of lines matched without human touch, days to close the period, days sales outstanding, and the share of receivables past sixty days. Reconciliation effort commonly falls by half or more; the residual is genuine discrepancy that always needed a person.

Where it does not pay: a few dozen documents a month will not recover the build. Neither will strictly regulated filings with fixed formats, where rule-based software is cheaper and more predictable. The division of labour is stable - deterministic rules for the strict parts, AI for the messy parts.

Operations and logistics

Forecasting demand and load is the oldest proven application and predates LLM entirely, but it creates value only where the forecast changes a decision: a purchase order, a shift roster, a route, a capacity commitment. Allocation assigns incoming jobs to crews, routes or machines by skill, location, current load and promised time. Deadline control watches orders whose intermediate milestones have slipped and raises them while there is still time to react, with the probable cause attached rather than a bare alert. Variance analysis reads the free-text exception records written by whoever was closest to the problem and clusters causes by supplier, shift, product or region - turning "we were late forty times this quarter" into "two suppliers account for sixty per cent of it."

Measure forecast error at the horizon at which decisions are actually made, then what it drives: stockouts, overtime, idle capacity, expedited shipping, on-time completion, the spread of utilisation across the team.

Where it does not pay: erratic demand with no usable history - a new product line, a business under two years old - gives a forecast nothing to learn from. Neither does a fixed-capacity situation where the decision would not change whatever the forecast said. Better information that cannot alter an action is an expensive comfort.

Documents and legal routine

This is where retrieval-based systems, RAG, do work nothing else does well. Search and answers over corporate documents - policies, regulations, contracts, technical documentation, records of past projects - with a citation to the specific clause, which is non-negotiable here: an answer without a source is unusable in a legal or compliance context. Contract comparison answers the question an editor's diff cannot: what changed against our standard template, and which changes carry consequence - liability, indemnity, termination, payment terms, jurisdiction. Term extraction builds a register of obligations across the whole portfolio: renewal and notice dates, indexation, penalties, exclusivity, volume commitments.

Measure the volume of questions escalated to the two or three people who know everything, the time a new employee needs to become independently productive, and review time on a returned redline. For many companies the extraction step is the first time their own obligations are visible in one place, and the common finding is a contract that auto-renewed because nobody was watching the notice window.

Where it does not pay: a document set small enough for one person to read needs no retrieval. An archive of unsorted scans without a text layer turns the project into an OCR and cataloguing exercise, which is the real cost and should be priced as one. Anything leaving the company as a legal position stays assisted, never autonomous.

People and hiring

Applications can be screened against the requirements that actually matter, producing a ranked shortlist with written justification a hiring manager can argue with. Role descriptions can be built from what the team actually does and the profiles of people already succeeding in the role, rather than from a template that yields the same generic posting as every competitor. Internal questions - leave, expenses, equipment, access, tooling - run on the same retrieval mechanism pointed at internal documentation. Measure time from application to first contact, the share of applicants reviewed within twenty-four hours, and ticket volume reaching HR and IT.

Where it does not pay: five hires a year is not volume. And a model must never make the rejection decision itself - in the EU that carries direct exposure under automated decision-making rules, and analytically it optimises for the wrong thing. The system ranks and explains; a person decides.

Leadership: the panel where the owner sees the business

The request behind most executive dashboards is not more charts. It is two things: a short set of numbers that genuinely drive this business - cash position, backlog, margin by line, on-time delivery, pipeline coverage - and an explanation layer that says why a number moved, in plain language, with drill-down to the underlying records.

The second part is what has been hard until recently. Assembling metrics is standard BI. Explaining a deviation - that the margin drop is concentrated in one region, driven by two customers on discounted terms after a supplier price rise - was a week of analyst work, requested only once the deviation was already large. Automating it changes the tempo of management, so the sensible metric is days between an event happening and someone acting on it.

Where it does not pay: when source systems disagree and revenue has two versions depending on where it is read. Then the project is data reconciliation, and calling it AI postpones the real work. It also fails when the owner already has the numbers and does not act on them - a tooling purchase covering a management problem.

How to tell where AI pays from where it is decoration

Four questions separate the two, and all four have to hold.

• Repeatability. Does the task have a recognisable shape - similar inputs, similar outputs, a describable notion of a good answer? A useful test: could you explain it in writing to a competent new hire? If your own team cannot articulate how the decision is made, a system will not infer it.
• Cost of an error. What happens when the output is wrong, and how easily is it reversed? A wrong sentence in a draft gets edited; a wrong payment, legal position or client message does not. High cost does not disqualify the task - it moves the design from automation to assistance, and moves the business case from headcount to the speed and quality of the person who stays in the loop.
• Volume. The build is a fixed cost; the saving accrues per event. The crude test takes ten minutes: hours per month on the task, times loaded hourly cost, times twelve, against build cost plus a year of running cost. If the annual saving does not clear the build within roughly a year, the task belongs on a different list - not because the technology failed, but because the arithmetic did.
• Data. Does the input already exist in machine-readable form, in one place, with an owner? This is where most stalled projects stall, and the honest answer to "why is this taking so long" is usually here. The model was never the constraint.

All four means a project. Three out of four means fixing the missing condition first, and it is almost always the data. Two or fewer means leaving the task alone for now, whatever the demo looked like.

Where this leaves a first project

The pattern that works is narrow: one function, one process, baseline measured before anything is built. Companies that report disappointment with AI are rarely the ones whose models underperformed. They are the ones that never wrote down what the process cost beforehand and therefore could not show that anything improved. What separates a return from an expense is a number recorded before the work starts, and the willingness to state in advance the condition under which the project would be judged not to have worked.

AKVANT Technologies is an AI consulting and engineering firm. We work with companies to find where AI pays off — and build only there.