Blog
/
AI implementation

AI implementation consultants: what they cost, what they actually do, and when you need FDEs instead

15 min read
Table of contents
Section oneSection two

Companies search for AI implementation consultants at the exact moment the story stops being fun: the pilot stalled, the board asked for a plan, or the team that was supposed to "figure out AI" figured out that it is a second job. The searches themselves quadrupled over the past year in our keyword data, which tracks the underlying reality: 95 percent of enterprise generative AI pilots produce no measurable P&L impact per MIT's Project NANDA, and RAND notes that by some estimates more than 80 percent of AI projects fail, twice the rate of ordinary IT projects. Failure at that scale creates a market for rescue, and consulting is the rescue industry.

This guide is the buyer's briefing: what implementation consultants actually deliver, what each tier of the market charges, the four questions that separate a good engagement from an expensive PDF, and the honest decision between renting expertise and building the in-house forward deployed engineering muscle that the best AI companies now consider standard equipment. We build software for FDE teams, so we have a stated view on the last question; the analysis before it stands on its own either way.

What AI implementation consultants actually do

"AI consultant" covers five distinct jobs, and engagements go wrong when buyers think they are purchasing one and the firm is selling another.

Strategy and readiness. Use-case discovery, feasibility, data-readiness assessment, and the roadmap deck. Genuinely valuable when leadership cannot rank its own opportunities; famously vulnerable to ending at the deck. RAND found 84 percent of its interviewees cited leadership-driven causes, wrong problem, undefined success, as the primary reason AI projects fail, and good strategy work exists to prevent exactly that.

Data engineering. Pipelines, cleanup, retrieval infrastructure. The unglamorous majority of real implementation effort, and the layer where underscoped engagements go to die.

Build and pilot. The POC: a working system against a scoped workflow. Most firms are competent here, which is precisely why the market is drowning in successful pilots and starving for production systems, a gap we dissect in why AI implementations die after the POC.

Integration into production. Wiring the working system into the real environment: the ERP, the permissions, the security review, the workflows. This is the scarcest skill on the list, the one that most resembles forward deployed engineering, and the line item to scrutinize hardest in any proposal.

Enablement. Training, change management, and adoption work. Decisive for internal-facing deployments; MIT's research found the divide between the 5 percent and everyone else is workflow adaptation, which is as much organizational as technical.

A serious engagement names which of the five it covers and which it explicitly does not. A vague engagement covers all five in the proposal and one and a half in practice.

What they cost, tier by tier

Rate guides from firms that track the market put 2026 pricing in consistent bands. AIDOLS' cost guide maps the tiers: independent consultants at roughly $150 to $350 an hour, boutique AI firms at $200 to $500, mid-tier firms at $300 to $600, Big Four at $400 to $800, and strategy-house partners at $500 to $1,000 and above. Leanware's guide puts mid-level hands-on rates at $150 to $300 an hour and seniors at $300 to $500, with project-based work from a few thousand dollars for a small assessment into six figures for enterprise transformations, and standard retainers at $5,000 to $15,000 a month.

Three buyer's notes on those numbers. First, the tier premium buys risk absorption and brand cover more than it buys better engineers; the person writing your integration code at a Big Four firm may bill at triple the independent who would do it better. Second, per-hour comparisons mislead: a $500-an-hour specialist who has built your exact integration five times is cheaper than a $150 generalist learning on your invoice. Third, the quote is rarely the cost. The visible engagement buys the build; the invisible tail, maintenance, changes, and the re-learning after the firm departs, is where the real total lands, and it lands on you.

For calibration, hold the rates against what the equivalent full-time talent costs: Indeed puts forward deployed engineer pay at $170,000 to over $200,000, and AI-lab total compensation runs $350,000 to $550,000 for mid-to-senior roles. A $200-an-hour consultant at full-time hours annualizes to roughly double the top of the Indeed band. Consultants are not expensive because consulting is a racket; they are expensive because you are renting optionality, and optionality has a price.

The three engagement shapes, and which one to buy

Underneath the rate cards, engagements come in three shapes, and matching the shape to your situation matters more than negotiating the rate.

The assessment sprint, two to four weeks. Senior judgment applied to your opportunity list, data readiness, and a sequenced roadmap. Priced at the strategy end of the hourly bands, or as a small fixed fee. Buy it when leadership genuinely cannot rank its options; skip it when you already know the workflow and just need it built, because re-discovering what you know at $400 an hour is the most common waste in the market. Exit criterion: a ranked list with one named first workflow and a success metric, not a maturity model.

The build sprint, six to twelve weeks. A scoped system against one workflow, from Leanware's mid-size band of $25,000 to $100,000 into six figures for production-grade work. AIDOLS puts proofs of concept at $50,000 to $250,000. This is the shape most buyers actually need first. The whole game is what "done" means: insist the definition includes integration into your real environment and the transfer test, or you have bought a demo with an invoice. Exit criterion: your operator ships a change without the firm on the call.

The embedded quarter, three-plus months. One or more of the firm's engineers working inside your team's rhythm, at the retainer or day-rate end of the pricing. This is consulting imitating the forward deployed model, and it is the honest shape when you need the muscle before you can hire it. Treat it explicitly as a bridge: the exit criterion is your own hire onboarded against artifacts the embed left in your custody, and the team structures they should slot into decided before they arrive.

Buyers who name the shape, the exit criterion, and the custody rule in the first call report a strange side effect: the weak firms stop returning emails, which is the cheapest vendor filter available.

Four questions that separate an engagement from a PDF

1. What lands, and where does it land? Insist on running systems in your infrastructure and your accounts as the deliverable, with strategy documents as the wrapper rather than the product. If the deliverables live in the firm's environment "for convenience," you are renting your own project.

2. Who operates it in month two? The system will need changes the week after go-live. Name the operator before signing: your team (then knowledge transfer is a scoped deliverable with dates), the firm (then you have bought a retainer; price it now), or nobody (then you have bought the sixth mechanism of the post-POC death spiral).

3. Where do credentials and code live during and after? The engagement will touch real systems with real secrets. The acceptable answer is your custody from day one: your repos or platform, your vault, access you can revoke at contract end. The common answer is the consultant's laptop, discovered at the security audit.

4. What does the firm keep? Pattern knowledge walks out at every contract end; that is the business model, not a betrayal. The question is whether anything compounds on your side: runbooks, reusable components, trained staff. Engagements that leave nothing behind but a working system leave you owning a system nobody understands, which is a liability with a go-live date behind it.

Red flags to catch at the proposal stage

Six tells, each cheap to check and expensive to ignore:

  1. Deliverables named as meetings. "Discovery workshops" and "alignment sessions" as line items mean the deliverable is the calendar. Workshops are how work happens, not what you bought.

  2. Environment access listed as "to be determined." If the proposal has not planned how the firm reaches your real systems, the engagement will spend its first month discovering your security process at your expense, and the pilot will quietly retreat to sample data.

  3. No named engineers. A partner sells the engagement, and the staffing appears later. Require the resumes doing the work, and the right to approve substitutions; the tier pricing above only means anything attached to specific humans.

  4. Model-selection theater. Proposals that spend pages comparing frontier models are marketing the easy 10 percent. The models are interchangeable relative to the integration work; a firm that leads with model choice is telling you where its comfort zone ends.

  5. Maintenance conspicuously absent. If the proposal ends at go-live, so does the system. Ask for the month-two rate card in the same document as the build quote; firms that resist are pricing the tail after you are captive.

  6. IP language that keeps the connectors. Some contracts grant the firm reusable rights over integration components built on your dime, or worse, deliver them as a service from the firm's infrastructure. Everything that touches your customers and credentials is yours, in your custody, full stop.

None of these makes a firm dishonest; each makes an engagement shaped for the firm's economics rather than your outcome. The fix is never confrontation, just specificity in the statement of work.

How to run the engagement, week by week

Buyers get the engagement they operate, not the one they signed. Four disciplines keep it pointed at production.

Week one produces access, not analysis. The measurable kickoff outcome is the firm's engineers working against your real systems under your custody rules: accounts provisioned, credentials vaulted on your side, repos or platform access granted. Every week spent on slides before access is a week the integration gap stays unexamined.

Demos run against reality, on a weekly cadence. The standing demo shows the system operating on your data in your environment, however small the increment. Sample-data demos are permitted exactly once, at kickoff. This single rule surfaces the integration gap in week two instead of month four, while the budget can still respond.

Custody checkpoints ride the invoices. At each billing milestone, verify the artifacts are where the contract says: code in your repos or platform, secrets in your vault, documentation in your systems. A five-minute check per invoice beats a forensic reconstruction at contract end.

The last deliverable is your team operating alone. Gate the final payment on a transfer test: your operator ships one small change end to end without the firm on the call. It is the same verification we prescribe for departing FDEs, because a consultant rolling off is a departure with an invoice attached.

And notice the mirror image: the best firms interview you as hard as you interview them. A consultant who asks about your success metric, your operator, and your custody setup before quoting is showing you their delivery model. One who asks only about budget and timeline is showing you theirs, too.

When consultants are the right answer

The honest cases are real. A bounded, one-time transformation with a defined end state, a migration, a first data platform, plays to consulting's strengths: pattern knowledge from a hundred prior projects, applied once, no permanent seat needed. First-time strategy, when leadership genuinely cannot rank its AI opportunities, is worth senior outside judgment. Regulated verticals sometimes justify firms that live in the compliance details. And surge capacity ahead of a hard date can be worth the premium when the alternative is a missed quarter.

What unifies the honest cases: the work ends. Consulting is the right shape for problems with finish lines. The corollary discipline: write the finish line into the engagement before it starts, because open-ended consulting is not a shape at all; it is in-house staffing at triple the rate with none of the retention.

When you need FDEs instead

The consultant question changes entirely when the work recurs, and for one class of company it recurs by definition: software vendors whose product needs custom integration at every serious customer. If each new deal generates environment-specific work, you do not have a project; you have a motion, and renting a motion is the most expensive way to own one.

Three structural reasons the in-house forward deployed engineer wins the recurring case. Knowledge compounds: the FDE's tenth customer build is faster and better because of the nine before it, while a firm's tenth engagement is priced like the first. Ownership persists: the code, the credentials, the context stay with people who still work for you next quarter. And economics invert at volume: consulting scales linearly with engagements, while an FDE team on a real operational platform gets cheaper per build as patterns accumulate. This is precisely why the market invented the role: Bloomberry counted 1,165 percent year-over-year growth in FDE postings as companies concluded the last mile was a permanent function, not a project. If you go this way, the operating manual is our guide to how to run an FDE team.

The two answers also hybridize well, with one rule: consultants for the first instance, FDEs for the pattern, and every deliverable lands in your custody regardless of who built it. Many strong teams use a boutique for the novel first integration in a new vertical, then productize the pattern internally.

Run the arithmetic on the recurring case and the crossover is not subtle. Take a vendor closing two integration-heavy deals a month. At the low end of Leanware's mid-size project band, around $25,000 per scoped build, that is on the order of $600,000 a year, with the twenty-fourth build priced like the first and the knowledge resetting at every contract boundary. The in-house version, one FDE at the top of Indeed's $170,000 to $200,000 band plus platform costs of $45 per bridge per month after the first two free bridges, comes in well below that in year one, and the gap widens every quarter as patterns accumulate and the marginal build gets faster. The numbers are illustrative rather than a quote, but the shape is not: rental pricing is flat, ownership pricing compounds downward, and the recurring case crosses the line fast.

The market itself is converging on the same conclusion from both sides. The consulting industry has begun hiring the role by name, with Deloitte running public requisitions for forward deployed engineers, which tells you enterprise clients are now asking firms for forward deployed delivery in those words. When the rental market starts stocking the thing you were deciding whether to own, the category question is settled; only the staffing question remains.

The custody rule that saves either path

Whichever way you staff it, the operational layer decides whether the work survives the staffing. This is the part we build for. On Archway, the platform for forward deployed engineering teams, every customer-specific build ships as a bridge: a serverless function connecting the product to one customer's systems, deployed in minutes with no core-product change. Credentials sit in an AES-256 vault with zero-access custody, so neither your engineer nor your consultant ever sees the values, which turns question three above into a checkbox. Every version is retained and owned by your organization. When a consultant rolls off, you reassign the bridges inside your org and leave a note: the bridges stay, the history stays, the successor takes over.

That property reframes the consultant decision. The risk of outside builders was never their competence; it was that their work lived in their world. Require the work to land as bridges in your account, and you get consulting's speed with ownership's permanence: two bridges are free, so the requirement costs your consultant nothing but an argument. The full software context for everything else in the stack lives in our guide to AI implementation software.

Five clauses to negotiate into the statement of work

Everything in this guide compresses into contract language, and the five clauses below cost nothing to ask for and reshape the engagement's incentives. Deliverable location: all code, configuration, and documentation land in buyer-owned repositories or platforms continuously, not at milestones, so contract termination at any point leaves you holding the current state. Credential handling: the firm's engineers access buyer and end-customer secrets only through the buyer's vault, never through local storage, with the buyer able to revoke unilaterally. The transfer test: final payment gates on the buyer's named operator shipping one change end to end without firm assistance, defined this concretely, because vague "knowledge transfer sessions" satisfy the letter and transfer nothing. Substitution approval: named engineers, with buyer approval required for staffing changes, which keeps the tier premium attached to the humans who justified it. The maintenance rate card: month-two-and-beyond pricing in the same document as the build quote, because negotiating it after go-live is negotiating from captivity.

Firms that accept all five without drama are telling you they deliver this way anyway, which is the strongest positive signal available at the proposal stage. Firms that fight the first three are telling you where the engagement would have hurt, at the only moment the information is free. Send the clauses with the RFP, not after the pitch.

The decision in five lines

One-time transformation with an end state: hire the consultant, land the deliverables in your custody. Recurring customer-specific work: build the FDE muscle, platform first. Cannot rank your opportunities: buy senior strategy by the week, not the quarter. Under deadline with neither: hybrid, consultant for instance one, FDE for the pattern. And in every branch: the code, the credentials, and the versions live where the staffing cannot take them. If that last line is not true of your current setup, fix it before you sign anything; the first two bridges are free, and it is a better use of a week than another vendor call.

Sources and notes

A long list of live bridges: approval rules, usage reports, data migrations, SLA alerts, reconciliation, provisioning and compliance exports

Deliver hyper-personalized software at scale

Start free

Product

ProductSecurityPricing

Legal

Terms of UsePrivacy PolicyPurchase Terms

© 2026 Archway. All rights reserved.