AI strategy: what you're buying and how to tell whether it worked
Four different services are sold under the name AI consulting. Which one you need first, what each actually leaves behind, and how to tell the difference between an engagement that will change something and one that will end in a deck.
In short
Decide which of four things you are buying before comparing anyone: strategy, architecture, delivery or governance. They have different deliverables and different right moments. An engagement that does not name which one it is defaults to strategy, because strategy is the cheapest to deliver and the hardest to be held to. One question separates useful work from decorative work: what number was measured before the work started. Impact scores without a baseline are opinions arranged in a table.
Who this is for
- A CEO, CTO or COO who needs to decide where to invest in AI and explain that decision to the board.
- A company that has run pilots and does not know which to move into production and which to stop.
- A company with proposals from several AI vendors that wants an independent view before it commits.
- An organisation with older systems and scattered data that needs to know what to prepare before it starts.
Four purchases, one name
| What | The question it answers | What it leaves behind | Right moment |
| Strategy | Which problems are worth tackling at all | A short ordered list, with each item paired with the measurement that would prove it worked | Only when you cannot name the task you want removed. Otherwise, skip it |
| Architecture | How a system that touches your data should be built, and where the boundaries are | Where data flows, what the model may see, what is deterministic, what is escalated | Before the first build, and again before the first system goes near customers |
| Delivery | Build it | Working software, and the obligation to run it | After the boundary above is written down. Before that, you are paying to discover it twice |
| Governance | What the organisation is allowed to do, and how that is evidenced | Versioned rules, an audit trail, and a named owner for each decision type | In parallel with architecture, not after it. Retrofitting it changes where data flows |
The last row is the one most often sequenced incorrectly. Governance is treated as a compliance step at the end, but what it actually constrains is architecture: whether data leaves the organisation, whether a decision has to be explainable, whether a model version must be recorded. Decide those things late, and the constraints arrive as a rebuild.
The question that separates good engagements from bad ones
One question, and it is not about technology: what number will be measured before anything starts?
The common deliverable in this market is a list of opportunities scored by impact and effort. It looks rigorous, and it is unfalsifiable. Nothing in it can be checked in six months, because nothing was measured at the beginning, so any outcome can be described as a success.
The alternative is unglamorous. Pick the task you want removed. Measure how long it takes today, how many people touch it, and how often it happens. Three numbers, collected in a week, without a consultant. Everything that follows is judged against them.
A second question worth asking: "What would make you tell us not to do this?" An engagement with no answer has only one possible conclusion, which means the conclusion was not produced by the work.
What you get
- Three baseline numbers for the chosen task: how long it takes today, how many people touch it, and how often it happens. Everything that follows is compared against them.
- The boundary, written down as decisions rather than features. What the system answers itself, what it hands off to a person, what the model must not see.
- A build or buy recommendation for each task: your own development, a ready-made product or a mix, including the risk of depending on a single vendor.
- A benefit and cost model with its assumptions written out. After the pilot, it is checked against the baseline rather than taken on trust.
- The governance constraints that affect architecture: where data may sit, which decisions must be explainable, what has to be logged.
- An order of work with the risk named at each step and a decision point: continue, change or stop.
What you will not get: a scored opportunity list before anything has been measured.
How the work runs
- A session with management: business goals, constraints, and what has already been tried.
- Interviews with unit heads, IT and the data team: where the manual work is, where decisions are slow, and where the data lives.
- Measuring the chosen tasks: the three baseline numbers. This step includes a technical review of the systems, the data and any pilots already running.
- Analysis: tasks are compared using the measured numbers, complexity and risk, and for each one we decide whether to build or buy.
- Presentation to management, revisions based on the comments, and handover.
If you want, we stay with you through the first pilot and into production. You can also deliver with another team: the result is written so any team can work from it.
What a governance deliverable actually contains
Governance is the row most often bought and least often specified. A useful deliverable is four artefacts. If an engagement produces a policy document instead of these, it produced reading material.
- A list of decision types, each marked allowed, allowed with a human, or not allowed. Not principles but named decisions: refuse an application, set a price, answer a customer about their account, draft an internal summary. The value is in the specificity, because that is what an engineer can implement and an auditor can check.
- Rules with versions. If a policy changed, a decision from last year has to be explainable under the policy that applied at the time. A system that stores only the current rules cannot justify its own history, and that is exactly what gets asked.
- A data map on one page: what leaves the organisation, to which service, under which agreement, and what is retained there. One page, not an appendix, because it has to be read by people who do not read appendices.
- A named owner for each decision type. Not a committee: the person who answers when that decision goes wrong.
The fourth is the test of whether the rest is real. A policy with no named owner describes an intention; the same policy with a name against each line is a commitment someone has accepted.
And a question that shows how an engagement will end: "Who maintains this after you leave?" Governance does not stay true on its own. Models change, vendors change terms, the business adds a use case. If no one inside owns the list, it is accurate for about a quarter.
Conversational AI
Some of the searches that bring people to this page are about conversational AI specifically, so here is a direct answer.
Any vendor can deliver a working assistant today; that is no longer the hard part. The hard part is the boundary, and it comes down to four decisions:
- What it answers from your own content, and what it must not answer at all. An assistant that was never taught to refuse will not decline; it will invent an answer, and do so confidently.
- What happens when it does not know. Handing over to a person has to be designed, including what that person receives. "I cannot help with that" is not a handover.
- How a wrong but confident answer is detected. Not by a user complaining, but by sampling real conversations against the sources, deliberately and on a schedule.
- Who is accountable for what it tells a customer. The legal side of this question comes before the technical one, and it should be settled in writing before launch, not after the first incident.
If the assistant has to answer from data inside your own systems rather than from your public pages, that requires a different architecture with different access control. More: integrations.
What the Israeli context adds
- Hebrew output needs a native signature. Review by a speaker is not enough; it takes sign-off by someone who writes the language professionally. Slightly wrong Hebrew from an assistant reads as careless, the team that built it does not notice, and customers see it from day one.
- Multilingual is the normal case, not an edge case. The same person may read Hebrew, work in English and receive content in Russian. Language belongs to the person and to each message, not to the whole system. A missing translation must be visible as missing, not quietly filled in with another language.
- Jurisdiction is a design input. Where data may sit, and whether a decision affecting a customer has to be explainable, are legal questions, and they determine architecture. A consultancy that leaves them for a later checklist has designed around assumptions it did not test.
Our principles
- We start from the business problem, not from the technology.
- One task working in production is worth more than several pilots that stayed demos.
- Every assumption in the benefit model is written down and can be checked. We do not promise a return.
- Every task has an owner on your side and a decision point, set in advance, where you decide whether to continue.
- A review is paid for on its own and does not depend on any development that may follow. So the conclusions "this does not need AI, it needs a process fixed" and "your volume is too low to pay for this" cost us nothing, and in the cases we have measured, they are the common ones.
Ways of working
- Remote. Interviews and working sessions by video, secure access to documents and data. Suits most of the measurement and analysis.
- On site. Workshops with management and unit heads at your offices, meetings with the teams doing the work. Suits cases where several units have to reach one decision.
- Hybrid. An opening workshop with you, remote work after that, and the final presentation to management in person. Often the right balance of depth and time.
Why us
- We have been building software systems since 2004: 150+ projects for clients in 14 countries. We have worked on AI systems since 2023.
- We build and run systems ourselves. Our flagship system has been in production since 2008, so our recommendations are tested against what happens after launch.
- The measurements, the model and the order of work belong to you, and any team can carry them forward.
Frequently asked questions
How long does the work take?
It depends on the size of the organisation and how many units take part. We set the scope in the first meeting and give a grounded timeline after it.
Do you recommend a particular model or vendor?
Only once the task is defined. The recommendation is based on the requirements for accuracy, security, data location and cost, and it includes a fallback in case the vendor changes terms.
We already have pilots. Is this still relevant?
Yes. Pilots give real numbers on accuracy, cost and usage. For those tasks, they are a ready baseline, and we start from them.
How is this different from the AI opportunity map on the site?
The opportunity map is a short questionnaire that gives a first direction. This work is built on interviews and on measuring your tasks, and the result adds the boundary, build or buy decisions, governance and an order of work.
Who needs to take part on our side?
A management representative to lead the process, the heads of the relevant units, and someone from IT or the data team. Each of them needs only a small amount of time: interviews and review of the results.
Do you also carry out the plan?
We can: we develop AI systems and take them into production. But the review is sold on its own and is not tied to development, and the result is written so that any team can carry it out.
SLAtech LTD