Generative AI Consulting Services: What Are You Paying For?

There are two products sold under this name and they have almost nothing in common. One ends with a document: a strategy, a roadmap, a prioritised list of opportunities. The other ends with something running in your business. Both are legitimate. Paying for the first while expecting the second is the most common way firms waste six figures on this category.

So the question to settle before you talk to anyone is which one you are buying, and whether the firm you are talking to can actually deliver the other.

Here is what matters most:

• Two products: an engagement that ends in a document, and one that ends in a working system. Decide which you need.
• If the engagement ends at a recommendation, the build gets quoted separately by people who did not do the analysis.
• Ask what the consultant has personally put into production. Strategy advice from people who have never shipped tends to be unbuildable.
• The deliverable to insist on is a prioritised list where each item names the data it needs and who owns it.
• Fixed-scope pilot first. Programme-level engagements before anything has shipped are how budgets disappear.

What the two products actually are

Advisory. Interviews across your business, an assessment of where the technology could apply, a prioritised roadmap, sometimes a governance framework and a training plan. Delivered as documents and presentations. This is what the large firms sell and it is not fraudulent: for a big organisation with genuine coordination problems across many business units, having one agreed picture is worth real money.

Delivery. Somebody scopes one specific process, builds a working system for it, puts it into production against real data, and hands it over with documentation and an alert path. Delivered as software that runs.

The mismatch happens because both are described in the same language. A proposal saying it will "identify and prioritise high-value generative AI opportunities and define an implementation roadmap" is advisory. Nothing in it will run. If you needed a working document pipeline, you have bought a plan to buy one.

Neither is better. But a mid-sized firm with one painful process almost always needs delivery and gets sold advisory, because advisory is easier to sell and carries no delivery risk.

The three ways these engagements fail

The recommendation is unbuildable. The roadmap names an opportunity that depends on data your firm does not hold in a usable form. This happens constantly because the analysis interviewed leaders about what they want rather than examining what the systems actually contain. A recommendation that begins with "consolidate the customer data" is a recommendation whose real project is a data migration nobody costed.

Nobody owns the output. The engagement ends, the document circulates, and it has no owner. Six months later a new initiative starts from scratch. Guard against this by requiring that every item in the deliverable names the person who would own it and the systems it touches. If a consultant cannot fill those two columns, the item is not a project, it is an aspiration.

The pilot was run on clean data. A demonstration built on a curated sample proves nothing about your business, because the thing that determines whether these systems work is how they handle the messy fifth-to-half of real cases. A pilot that has not met a photograph of a crumpled document, a record with a mismatched name, or a statement covering the wrong period has not been tested.

The questions that separate the two

"What have you personally put into production, and what broke?" The most useful question available. Anybody who has run these systems has failure stories with specifics. We have ours: a workflow of ours wrote into a spreadsheet tab that did not exist and reported every run as successful for six weeks before anyone noticed no data had arrived. Another logged 11,458 outbound messages as sent when every one had died on a billing error at the far end. Both were configuration oversights rather than exotic faults. A consultant with no equivalent story has either not shipped or is managing the conversation.

"Does this engagement end in a document or in something running?" Ask it plainly and get it in writing. If the answer is a document, ask who builds afterwards and whether they were part of the analysis.

"What data does your top recommendation depend on, and have you looked at it?" Not asked about. Looked at. There is a large difference between a leader saying the firm has good customer data and somebody opening the system and finding three conflicting addresses.

"What is the smallest version of this that could run in production in six weeks?" A good answer exists for almost every real opportunity. A consultant who cannot produce one is either scoping a genuinely large programme, which is sometimes right, or has not thought concretely enough.

What a useful deliverable looks like

If you are buying advisory, specify the format rather than accepting theirs. A deliverable worth paying for has, for each opportunity: the process it touches, how often that process runs, the mechanical hours it currently consumes, the data it depends on and the state that data is actually in, the named owner, the smallest production version, and what a wrong answer would cost.

That table is buildable from. A maturity model and a two-by-two grid are not.

Insist also on the arithmetic being shown rather than asserted. For each opportunity: runs per month, mechanical minutes per run, fully loaded hourly cost. Sixty runs a month at forty-five mechanical minutes is forty-five hours, around eighteen hundred dollars a month at forty dollars an hour. That is a number you can act on. A statement that a process could be "up to seventy percent more efficient" is not.

How we would sequence it

One process, scoped and built to production, before any programme-level engagement. That gives you three things no strategy document can: a real exception rate for your own data, a working system somebody depends on, and an honest cost per unit of work at your actual volume.

With those three numbers, the second project scopes itself and you no longer need a roadmap to decide it. Without them, a roadmap is a set of guesses ranked by confidence rather than evidence.

One framing worth holding to throughout: these systems should sit alongside your people. They read, gather, route and draft. Your team keeps judgement, review and the client relationship. Engagements sold on headcount reduction require cooperation from exactly the people being displaced, and they stall for organisational reasons long before technical ones.

FAQ

What is the difference between AI strategy consulting and AI implementation?

Strategy consulting produces analysis and a plan; implementation produces a working system. The practical consequence is who carries the risk: a strategy engagement is complete when the document is accepted, regardless of whether anything can be built from it, while an implementation engagement is only complete when something runs against real data. Firms frequently buy the first believing it includes the second.

Should we hire a large consultancy or a specialist?

It depends which product you need. Large firms are genuinely better at coordination problems across many business units, governance frameworks, and giving a board one agreed picture. Specialists are generally better at getting one process working, cost materially less, and are more likely to have built the thing they are recommending. For a firm under a few hundred people with one painful process, the specialist is usually the right purchase.

How much should a first engagement cost?

Rather than benchmarking a rate, size the process first: runs per month, mechanical minutes per run, fully loaded hourly cost. That gives you the annual value at stake, and a first engagement that costs a large multiple of it is not worth doing regardless of how reasonable the day rate looks. Insist on a fixed scope with a defined deliverable, because open-ended discovery is where these budgets go.

What if the consultant recommends a platform purchase?

Ask what they would do without it, and ask whether they have a commercial relationship with the vendor. Platform licences make sense at genuine enterprise scale where procurement and vendor governance are part of the requirement. Below that, the licence is often the most expensive and least useful line in the proposal, and the same outcome is reachable with tools you already pay for.

Key takeaways

• Generative AI consulting splits into engagements that end in a document and engagements that end in a running system.
• A recommendation is only actionable if it names the data it depends on and the person who would own it.
• Pilots run on curated samples prove nothing, because real performance is decided by the messy minority of cases.
• Ask what the consultant has personally put into production and what broke. Specifics indicate real delivery experience.
• Build one process to production before commissioning a programme-level roadmap.

If you have a proposal in front of you and want a second read on whether it ends in a document or in something running, we will go through it with you.

Book a free strategy call

Related reading: do you need an AI automation consultant · AI readiness assessment · AI agents for business

Services: business process automation consulting · AI readiness assessment