How to know where to actually invest in artificial intelligence

Every week we meet companies that ask the same question:

“We want to introduce artificial intelligence. Where do we start?”

The question sounds reasonable. ChatGPT is on the board table. A competitor launched an assistant. A vendor is pitching an agent. An executive saw a demo and asked for a plan.

That question already contains a mistake.

It is not a mistake of curiosity. It is a mistake of sequence. Teams start from the tool and look for a problem to attach it to. The order is the opposite: first understand where the company is losing time, money and productivity. Then decide whether the answer is an integration, a workflow, a model or an AI agent.

Artificial intelligence is not the starting point. It is the final multiplier of a strategy built on processes.

This article is an operating guide for CEOs, COOs, CIOs, CTOs, innovation managers and founders in companies of 50 to 1,000 people. It is not a technology catalogue. It is a method for knowing where to invest, before spending on a chatbot, a Copilot or a RAG project.

The question to start from is not “where can we use AI?”. It is: “which processes consume the most labour hours, generate the most errors and block the most value?”

Why so many companies start in the wrong place

The market sells products. Products have names. Names become projects.

In the last two years the internal conversation has changed. People used to talk about digital transformation in a broad sense. Today, in many meetings, they talk about chatbots, agents, Copilot, RAG, automations. All of these are real objects. All of them are useful in some contexts. Almost none of them, on their own, tell you where the company is losing margin.

This is how it happens.

Customer service receives thousands of emails. Someone proposes a site chatbot. Sales spends hours updating the CRM. Someone proposes a Copilot. Purchasing classifies PDFs. Someone proposes RAG over documents. IT receives three different requests, three proofs of concept, three vendors.

Nobody has measured the hours yet.

In the State of AI 2025 McKinsey found that 88% of organisations use AI in at least one business function, up from 78% a year earlier. Only 39% attribute any enterprise-level EBIT impact to AI. Among those who do, most say it accounts for less than 5% of EBIT. A small group, around 6%, stands out as high performers: not because they have more tools, but because they redesign workflows around AI instead of bolting a model onto unchanged processes.

The gap between adoption and value is not a technology mystery. It is a problem of choice.

Gartner, in a July 2024 forecast, estimated that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. Reasons: data quality, rising costs, weak risk controls, unclear business value. Rita Sallam, a Gartner analyst, put it plainly: after the hype, executives want a return. Many organisations struggle to prove one.

The Microsoft Work Trend Index described another split: knowledge workers already use AI tools, often on their own initiative, while the organisation lags on processes, permissions and measurement. The Stanford AI Index keeps documenting growth in capability and investment. All of this increases the pressure to “do something”. It does not say what to do first.

Four recurring traps

Starting from the vendor catalogue. If the offer is a chatbot, the problem becomes “we need a chatbot”. If the offer is an agent, every process becomes a candidate for an agent.

Confusing visibility with value. An assistant on the website is visible. An integration that removes copy-paste between ERP and CRM is not. In many companies the second is worth more.

Mistaking a proof of concept for a strategy. A prototype that classifies ten emails is not a roadmap. It is an experiment. Without volumes, exceptions, integrations and ownership, it remains a video for the board.

Automating the wrong process. We covered this when most business problems are still solved without AI and when choosing which processes are worth automating. Technology comes after the map.

The market is not acting in bad faith. It sells what it knows how to produce. The job of the people who run the company is different: do not buy a solution before you have named the problem in hours, euros and risk.

AI is not a technology project

Treating artificial intelligence as an IT project is convenient. It has a budget, a vendor, a go-live date. It is also insufficient.

An AI project that touches real work involves processes, people, organisation, data and governance. IT is necessary. It is not enough.

Processes

If the flow is confused, AI speeds up the confusion. If three teams run the same activity in three ways, a model will learn three ways, or one, and the other two will keep working by hand. Before a model you need an honest description: who does what, in which systems, with which exceptions.

People

The people who currently “hold the systems together” with copy-paste, phone calls and personal memory are not an obstacle. They are the primary source of knowledge. Without them, mapping is an org chart, not a process. The Work Trend Index confirms this in another way: people adopt AI before the company does. If you do not involve them, they will use ungoverned tools. If you involve them only at go-live, they will resist.

Organisation

Who owns the process after automation? Who decides when output is acceptable? Who pays for maintenance? Without ownership, the project returns to IT as tickets. AI Transformation is not an isolated centre of excellence. It is a line responsibility, with support from data and technology.

Data

An agent that cannot see the order in the ERP, the current policy and the open ticket is not intelligent: it is blind. Before agents you need context: we called it a Company Brain. Dirty data, duplicate records, unversioned documents: Gartner is right twice here. Without ready data, the proof of concept dies in production.

Governance

Permissions, audit, traceability, privacy, what happens if the system is wrong. This is not bureaucracy for its own sake. It is what makes innovation fundable. A CFO will not sign an investment they cannot explain if it fails on an order, a contract or a customer.

AI in the company is a transformation project. Technology is one lever. It is not the only lever, and it is almost never the first.

Anyone looking only for a software house will get a deliverable. Anyone looking for an AI Transformation partner gets a work order: what to change, in what sequence, with what expected return, and which technology is the least complex that solves the problem. That is the same logic as our digital consulting: real stack, real constraints, no slides for their own sake.

The right question to ask

“Where can we use AI?” produces a list of use cases. Use-case lists fill workshops. They rarely fill the P&L.

“Which processes consume the most labour hours?” produces a ranking. A ranking can be debated, measured and cut.

The difference is not semantic. It changes three things.

It changes the perimeter. You do not start from customer service because “language is used a lot there”. You start from the process that absorbs the most repetitive time, wherever it sits: finance, operations, sales, back office, quality.

It changes the success criterion. Not “we launched an agent”. But: hours taken out of low-value work, fewer errors, cycle time, cost per case.

It changes the technology. If the process is a sequence of rules, a workflow and system integration may be enough. If you need to interpret text or documents, AI comes in. If you need to reason and use tools, you can talk about agents. We drew that line in classic workflows versus agentic automation.

A small example.

An order desk receives requests by email, PDF and portal. Today a person reads, looks up the customer, checks availability, enters the order, updates the CRM, writes to the customer. Ten minutes. Two hundred times a week. That is more than 1,700 hours a year, before counting errors, rush jobs and double entry.

The question “where do we use AI?” suggests an assistant that replies to the customer. The question “where do we lose hours?” suggests measuring each step. Often the bottleneck is not the reply. It is ERP entry, master-data reconciliation, the status hand-off. AI can read the PDF. Integration writes the order. A person handles exceptions.

Same problem. Two different investments. Two different ROIs.

How to find the best opportunities

You do not need a two-year programme to get a first map. You need a repeatable method, applied to a real perimeter: one function, one product line, one end-to-end flow (order, complaint, accounts payable).

1. Map the process as it happens, not as it is drawn

Ask the people who run it to describe last week, not the official procedure. Note the systems they open, the spreadsheets, the exceptions, the “I ask Maria”. The gap between the drawn process and the lived process is often the real project.

2. Measure volumes

How many executions a month? Seasonality? Peaks? An eight-minute activity done 20 times a year is not a candidate. The same activity done 800 times a month is. Without volume, every saving estimate is a wish.

3. Measure time

Average time per execution. Waiting time between steps. Rework after an error. In many companies waiting time is worth more than working time: a case stuck for three days between two offices is not fixed by a chatbot.

4. Estimate operating cost

Hours × fully loaded cost (not salary alone). Include on-costs, tools, coordination. A “market” hourly cost between €28 and €45 is an order of magnitude for back-office roles in Italy, not an accounting figure. It is there to compare opportunities, not to close the books.

5. Count errors and their impact

Wrong data in the ERP, duplicate order, late invoice, customer calling back, a compliance file. Error has a direct cost and a trust cost. Processes with high error impact rise in the ranking even when volume is average.

6. Isolate repetitive work and bottlenecks

Copy-paste, classification, data entry, reconciliation, hunting for the right document version, the weekly report rebuilt by hand. These are the same signals we use in process automation.

A simple discovery table helps you keep the thread.

Score volume, time, standardisation, error impact and fragmentation from 1 to 5. Add them up. It is not science. It is a filter against fashion.

Box · Automation Opportunity Score. Five dimensions, each from 1 to 5. Processes in the top right (high impact, manageable complexity) are the first project candidates. Processes in the top left (high impact, high complexity) need a deeper assessment, not an improvised proof of concept. Scoring detail is in the guide on which processes to automate.

How to calculate ROI without telling yourself stories

An AI business case that does not start from hours is an exercise in optimism.

The chain is simple.

Hours today → share that can be automated → hours recovered → hourly cost → annual saving → investment (build, integration, licences, change, maintenance) → payback → ROI.

Formulas, in plain language:

• Annual saving ≈ recoverable weekly hours × 46 weeks × hourly cost
• Payback in months ≈ investment / (annual saving / 12)
• 12-month ROI ≈ (annual saving − annual run cost) / investment

Three caveats.

The automatable share is not 100%. On many real operations it sits between 35% and 70% of repetitive time, because exceptions, checks and relationships remain. Promising 100% is the fastest way to burn trust.

Saving is not immediate cash. It is freed capacity. It becomes cash if you cut overtime, absorb growth without hiring, or reduce errors that cost money. Say that honestly to the CFO. McKinsey’s 2025 survey shows enterprise EBIT impact comes later, and for few. Function-level value (cost reductions of 10-20% in some engineering, IT and manufacturing cases; revenue lifts above 10% in some marketing and product cases) is more common than an effect on the consolidated P&L.

Investment is not only the model. It is integration, data quality, training, monitoring, human fallback. Gartner has stressed that generative AI costs are not as linear as classic software. A cheap proof of concept can become an expensive exercise in production.

A numerical example (a scenario, not a client)

Imagine a back-office team in a mid-market manufacturing company. It reports 40 hours a week of repetitive work on the order flow: reading requests, ERP entry, CRM updates, confirmation emails.

Conservative assumptions:

• real repetitive work band: 35-45 hours/week
• automatable share in year one: 40%
• fully loaded hourly cost: €35
• weeks: 46

Hours recovered per year ≈ 40 × 0.40 × 46 = 736 hours.

Indicative saving ≈ 736 × 35 ≈ €25,800 a year.

If the investment to connect email/PDF, ERP and CRM, with controls and an operator on exceptions, is €40,000 in year one (project + launch) and €12,000 of run cost from year two, payback is in the order of 18-24 months. It is not magic. It is an order of magnitude.

If the same management had spent €40,000 on a first-line website chatbot, with 200 conversations a month of which 40 are truly resolved without an agent, the saving would be a different story: smaller, more visible, weaker on the P&L.

Same budget. Two different starting questions.

Box · what not to put in the ROI. Do not attribute extra revenue “because we are now innovative”. Do not treat manager meeting time as a software cost. Do not use the company’s highest hourly rate if administrative profiles run the process. The business case has to survive a meeting with controlling.

Why many companies implement the wrong project

Take a hypothetical case. No names. The mechanics will feel familiar to anyone who works with production, distribution, B2B.

A manufacturing company, 180 people, Italy. Customer service handles about 25,000 emails a year (a little over 2,000 a month), plus the phone. Management has seen conversational-assistant demos. The request to IT is: “let’s do a chatbot, including an internal one, so we can scale.”

Analysis of the flow tells another story.

Email is only the surface. Underneath: classifying the request type; looking up the customer in three master-data sources; opening the ERP for availability and delivery notes; copying data into the CRM; creating a ticket; sometimes archiving a complaint or certificate PDF; updating order status; replying.

The time is not in the “conversation”. It is in the handoff between systems and in the paper.

A website chatbot might close a fraction of FAQs (“where is my order?”) if tracking data is already exposed reliably. Often it is not. The bot becomes a new channel that still queries an operator.

The higher value, in the same perimeter, sits elsewhere:

• classifying and routing documents (complaints, certificates, confirmations);
• updating ERP and CRM without double entry;
• closing the order cycle with consistent statuses across warehouse, sales and support.

The chatbot, if it is needed, comes later: when data is exposed, FAQs are real, and the conversational channel is not the only way to cover an integration hole. It is the same lesson as accounts payable: first extract and check documents, then orchestrate.

The wrong project is not “the chatbot”. It is the chatbot as a first step, chosen because it is visible, not because it is the constraint.

PwC and Deloitte, in their transformation reports, have repeated an adjacent point for years: value arrives when the way of working changes, not when you add an interface. IBM, on adoption indexes, has documented more than once the missing jump from experiment to scale. You do not need another statistic to see it in a company: ask how many proofs of concept from the last 18 months are still on on Monday morning.

The idea of an AI Transformation Assessment

At this point the argument asks for a tool.

Not another proof of concept. An AI Transformation Assessment: guided work to understand where artificial intelligence, automation and integration can create economic value, in what order, with what risk.

It is not a sales page under another name. It is the natural next step of what any operations director would do with two weeks and the right data: map, measure, choose, plan.

The output is not a slide that says “we are AI-ready”. The output is a package you can decide on.

AI Opportunity Map. Candidate processes, with volume, time, systems, constraints. A map, not a catalogue of slogans.

Business case. Expected saving, assumptions, what is not included, what depends on data.

Priority matrix. Impact vs effort vs risk. What to do now, what to do later, what not to do.

Roadmap. 30, 60, 90 days and beyond: deeper discovery, integration, optional AI, measurement. Not three tools to buy on Monday.

ROI analysis. Payback, run costs, indicators to follow after go-live. If a metric does not drive a decision, it does not enter the dashboard: we wrote about that in BI and decisions.

An assessment like this puts the work where it belongs: before choosing a model, before the vendor RFI, before the “AI 2026” budget is spent by spreading it thin.

Whoever runs it should not only be someone who sells licences. They should be able to read a process, an ERP, a CRM and a ticket queue. In Italy this craft sits between transformation consulting, system integration and Agentic Engineering. It rarely sits in only one of those boxes.

The Syncronika framework: six phases

At Syncronika we organise this path in six phases. The English names are method labels. The work is done on your systems.

1. Discover

Watch real work. Short interviews, shadowing, system list, main flows. Goal: a shared photograph, not a punitive audit. This is where “ghost processes” appear: spreadsheets, personal mailboxes, WhatsApp groups that keep the operation alive.

2. Measure

Volumes, times, errors, queues. Few numbers, good ones. If the data does not exist, sample a week. An honest sample beats an invented KPI. Without measurement there is no priority, only opinion.

3. Identify

Separate what is repetitive and deterministic from what needs interpretation or judgement. This is where you decide whether the lever is integration, workflow, AI on documents or text, or an agent with tools. Identify also includes removing: processes that should no longer exist.

4. Prioritize

Impact / effort / risk matrix. You pick one or two interventions for the first cycle. Not ten. Ten parallel initiatives are the classic way to finish none. McKinsey’s high performers do not win because they run more experiments. They win because they tie AI to a workflow change and to growth objectives, not only to an assistant.

5. Transform

Build the simplest thing that proves value on the chosen process. Integration, rule, document extraction, optional model, optional agent. With permissions, logs, fallback. Transformation is on the process, not on the press release.

6. Optimize

Measure afterwards. Compare the business-case hypotheses. Correct exceptions, thresholds, routing. Only then extend. Optimize is the opposite of “we launched, let’s move to the next use case”.

This cycle is deliberately unglamorous. Unglamorous, in this craft, is a compliment. It means you can repeat it next year on another flow without starting again from a vendor demo.

A fast version, on your own, is in the Process & AI Diagnostic: a few questions on context, processes, systems and readiness. It does not replace a field assessment. It helps you arrive at the conversation with a first reading of bottlenecks and opportunities.

Closing

The companies that will win in the next years will not be the ones with the most AI agents.

They will be the ones that identified, earlier than others, the processes with the greatest transformation potential, and had the discipline not to start from the tool.

Artificial intelligence will remain a multiplier. It multiplies what it finds. If it finds a dirty flow, it multiplies dirt, cost and exceptions. If it finds a measured, integrated process, with ownership and data, it multiplies productivity.

The first project should not be a chatbot. It should be a map. Then a number. Then a choice. Then, only then, a technology.

If your company is considering how to introduce artificial intelligence, the starting point is not choosing a tool. It is understanding where it can create the most value.

An AI Transformation Assessment identifies the opportunities with the best economic return and builds a concrete roadmap before you invest in any technology.

You can start with a guided reading in a few minutes, or tell us about a real process. In both cases you start from hours, not from the model.

FAQ

What is an AI Transformation Assessment?

It is a structured analysis of processes, volumes, costs, data and systems to decide where automation, integration and artificial intelligence create value, in what order and with what risk. It produces an opportunity map, a business case, a priority matrix, a roadmap and an ROI estimate. It is not the purchase of a model.

Are AI consulting and an assessment the same thing?

No. AI consulting can include strategy, architecture, build and governance. The assessment is the step before: understanding where it is worth moving. Without an assessment, consulting risks becoming the implementation of the most discussed tool of the quarter.

How long does it take?

A first reading (diagnostic) takes a few minutes. An assessment on a real perimeter, with interviews and numbers, is measured in weeks, not quarters. If you are offered three months before anyone names a process, you are buying a programme, not a decision.

Do we necessarily need an AI agent?

No. In many companies the first return sits in process automation and integration. An agent comes in when the process needs interpretation, tool use and a level of autonomy a workflow does not cover. Prove value on the flow first, then raise the degree of autonomy.

How do you measure AI ROI?

You start from hours (or errors, or cycle time), estimate the automatable share, convert to saving, compare with investment and run costs. You declare what is not included. You check after go-live. If ROI exists only on a pre-project slide, it is not ROI.

How is this different from a digital transformation project?

Digital transformation is the container: systems, data, channels, organisation. An AI Transformation Assessment is a way not to let AI become a separate chapter, detached from processes. It is digital transformation with an explicit economic criterion for where artificial intelligence multiplies, and where a rule is enough.

Where can a CEO or COO start on Monday morning?

Three things. Ask one function for the three processes that steal the most hours. Have them write volume, systems and errors for each. Do not approve a proof of concept that does not name those three variables. If you want a structure already prepared, use the diagnostic and then go deeper with people who know both the processes and the stack.