How to Choose an AI Consulting Firm: Seven Checks That Expose the Deck-Builders
Choosing an AI consulting firm takes seven verifiable checks, and delivery evidence exposes the deck-builders before a single reference call. Tender due diligence beats rapport, because delivery evidence outlasts a good first impression once the contract needs signing. The firm that wins every pitch on rapport and ships a slide deck is the firm these checks exist to expose, and a tender with no governance evidence or a demo that never survives month 12 both trace back to a selection made on chemistry rather than proof. Run all seven checks before the first meeting: a firm that fails any two of them is out, and a firm that trips a single red flag is out immediately.
Synthesised.ai publishes these checks knowing its own engagements are measured against them, because a guide worth citing sometimes disqualifies the firm that wrote it.
How to Choose an AI Consulting Firm: The Seven Checks
Seven checks qualify an AI consulting firm: delivery evidence, data-governance capability, pricing transparency, team composition, measurement discipline, a production record, and exit and handover terms, and each one separates an implementer from a deck-builder before a single reference call happens. Run every check against the shortlist before the first meeting, because a firm that clears the pitch on rapport still has to clear the matrix on evidence.
| # | Check | Industry Baseline | Expert Benchmark | Verification Method |
|---|---|---|---|---|
| 1 | Delivery evidence | Logo wall | Named, dated, referee-checkable systems | Referee call script |
| 2 | Data-governance capability | Privacy policy | Audit-ready governance artefacts | Artefact review |
| 3 | Pricing transparency | Rate card on request | Scoped, itemised quote | Quote comparison |
| 4 | Team composition | Partner pitch | Named delivery team in the statement of work | Statement-of-work clause |
| 5 | Measurement discipline | Post-hoc claims | Baselines defined pre-build | Method walkthrough |
| 6 | Production record | Demos shown | Systems running past month 12 | Uptime or audit evidence |
| 7 | Exit and handover | Silent on the topic | Transfer artefacts contracted upfront | Contract clause |
Run all seven checks before the first meeting, because a firm that fails any two of them is out, and a firm that trips a single red flag is out immediately.
By some estimates cited in RAND Corporation’s research on the root causes of AI project failure (2024; an interview study of 65 practitioners), more than 80% of AI projects fail, roughly twice the rate of non-AI IT projects, a figure RAND draws from earlier sources rather than calculating itself, which is part of why a firm is better qualified on evidence of delivery than on the strength of a pitch.
What Criteria Define a Quality AI Consulting Firm?
A quality AI consulting firm meets seven testable criteria: delivery evidence, data-governance capability, pricing transparency, a named delivery team, measurement discipline, a production record past month 12, and contracted handover. Each criterion carries its own verification method, so a claim the firm makes on the page becomes a claim the buyer checks in the room, rather than an adjective the buyer takes on trust.
How Many of the Seven Checks Must a Firm Pass?
The seven checks run a two-strike screen: a firm may miss one and stay in contention, but failing any two is out. A firm that trips a single red flag, covered next, is out immediately regardless of how many checks it otherwise clears. The checks run a two-strike screen, the red flags run a one-strike screen, and both gates apply together at the procurement stage.
How Is the AI Consulting Service Market Structured?
In this guide’s reading, the AI consulting service market splits into four supplier classes, and the seven checks apply to every one of them: Big 4 and global advisory practices, global systems integrators, independent engineering-led boutiques, and product vendors that sell consulting around their own platform. The class label says little about delivery, so the screen runs the same way whichever class a firm belongs to. The head-to-head on Big 4 AI consulting versus a boutique AI agency covers the first and third classes in depth, and the shortlist of the best AI consulting firms applies evidence tests across the wider field. A buyer whose use case centres on model development, forecasting or data science is choosing an AI and machine learning consulting firm, and AI and machine learning consulting sets out what that engagement covers.
Why References and Rapport Mislead Buyers
References and rapport measure how a firm sells, and the seven checks above measure how it delivers. Qualifying an AI consulting firm mirrors tender due diligence: referees, delivery records and priced scopes are checkable, and chemistry is not. Seven verifiable checks separate implementers from deck-builders before a single reference call, because a reference the seller arranges is pre-selected to say yes.
A hand-picked reference proves the firm produced one happy client, not that it delivers in production, because the referee is selected to pass before the call ever happens. The check that actually matters is a system running past month 12, not a warm phone call the seller arranged.
MIT NANDA’s The GenAI Divide: State of AI in Business 2025 report, as reported by Fortune on 18 August 2025, found that only about 5% of AI pilot programmes achieved rapid revenue acceleration, while the vast majority stalled with little-to-no measurable P&L impact; it measured P&L impact rather than the predictive value of vendor references, so Synthesised.ai reads it as a reason to weight production evidence over a demo rather than as proof that references predict nothing. A reference the seller arranges is pre-filtered for a positive answer, and the buyer never hears from the engagement that stalled; the counter is a referee call script the buyer controls, asking for named, dated systems and the outcome at month 12, not the contact the seller hand-picked. Rapport tells the buyer how the working relationship might feel, which is different from the delivery outcome; the seven checks are what isolate delivery from feeling.
Why Do Hand-Picked References Pass Every Firm?
Hand-picked references tend to pass a firm because the seller selects the referee, and a selected referee is likely to answer in the seller’s favour. The stalled engagement never gets offered as a reference, so the fix is a buyer-controlled call script that asks for the three most recent engagements, not the three the seller volunteers, and names the system running at month 12.
What Does Rapport Actually Predict?
Rapport indicates how a project meeting feels, which is a weak guide to whether the system survives production; chemistry and delivery discipline are largely separate, though RAND notes that communication and trust can also contribute to project success. Firms sell on rapport precisely because it is easier to demonstrate in a single meeting than a governance evidence pack or a system that has run past month 12; the root causes behind stalled AI implementations trace the failure a rapport-led selection quietly sets up.
Red Flags That Disqualify a Firm Early
Three findings end the evaluation immediately: no production system past month 12, no governance artefacts to show, and no willingness to name the delivery team. Each one is a one-strike disqualifier, not a point deducted from an otherwise strong pitch.
Gartner forecast in its 29 July 2024 press release that at least 30% of generative-AI projects will be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs and unclear value, and later reported (January 2026) that at least 50% had been. The share of companies abandoning the majority of their AI initiatives before they reach production rose from 17% to 42% year on year, and organisations on average report that 46% of AI projects are scrapped between proof of concept and broad adoption, according to S&P Global Market Intelligence’s 2025 Voice of the Enterprise: AI & Machine Learning survey of 1,006 respondents. Those figures evidence implementation and scaling risk, the reality behind the “nothing in production” red flag, rather than proving that demo-led selection caused those outcomes.
Red Flag 1: Nothing Running Past Month 12
A firm with no system running past month 12 has a demo track record, not a production one, and the gap disqualifies it on the spot. The verification method is uptime or audit evidence for a named, dated system, not a screen recording or a staged environment; AI Adoption Statistics quantifies how many AI projects never reach that production month 12 in the first place.
Red Flag 2: Governance Evidence Promised ‘Post-Engagement’
A firm that promises governance artefacts “after the engagement” has none of them now, and a governance pack produced after the build is a document, not a control. The verification method is an artefact review before signature: a control coverage table, data-flow diagrams and consent maps that already exist, not a promise of ones to come.
Red Flag 3: The Delivery Team Stays Anonymous Until Signature
A firm that names a partner in the pitch but not the delivery team in the statement of work has swapped the people who sell for the people who build. The verification method is a statement-of-work clause naming the delivery team, so the engineers in the pitch room are the same engineers on the contract.
What Single Finding Ends the Evaluation Immediately?
Any one of the three red flags ends the evaluation immediately: no production system past month 12, no governance artefacts to show, or no named delivery team in the statement of work. The red-flag screen runs one-strike, unlike the seven checks above, which tolerate a single miss before the two-strike rule applies.
The Compliance Checklist: Non-Negotiable Requirements
Documented alignment with the frameworks a procurement team verifies is strong due-diligence evidence, and where a tender, contract or procurement policy requires specific documents it becomes a requirement; the checklist below names a recommended document per requirement. A framework-alignment claim gets verified against a named standard or law, not taken on the strength of a slide.
| Requirement | Recommended Documentation | Framework or Law | Who Verifies | Timing |
|---|---|---|---|---|
| AI management alignment | Control coverage table | ISO/IEC 42001 | Procurement | Pre-signature |
| Risk function mapping | Function alignment table | NIST AI RMF | Risk owner | Pre-signature |
| SG deployment readiness | Framework alignment statement | IMDA AI Verify | Procurement | Scope-dependent |
| Data handling | Purpose maps, consent maps | PDPA | Legal or IT | Pre-signature |
| AU guardrail alignment | Guardrail mapping | DISR Guidance for AI Adoption | Risk owner | Scope-dependent |
ISO/IEC 42001 specifies an AI management system, and the control coverage table maps a firm’s controls to its clauses. The NIST AI Risk Management Framework organises the same accountability into GOVERN, MAP, MEASURE and MANAGE, and the function alignment table maps a firm’s risk function to that core. Singapore’s Personal Data Protection Act, with the PDPC’s AI Advisory Guidelines issued 1 March 2024, governs lawful data handling, and purpose maps and consent maps are the evidence of it. Every requirement above names the document a procurement panel asks to see, not synthesised.ai’s own status against it. The checklist tells the buyer what to demand of any firm.
What Certifications Does an AI Consulting Firm Need?
These are recommended due-diligence evidence rather than universal requirements: ISO/IEC 42001 is a certifiable international management-system standard, the NIST AI Risk Management Framework and IMDA AI Verify are voluntary frameworks (NIST states the AI RMF is intended for voluntary use; NUS Law’s TRAIL centre describes AI Verify as a voluntary self-assessment framework), and the Singapore PDPA is binding law where personal data is processed. A buyer asks for documented alignment to the ones its own tender, contract or procurement policy makes relevant, and verifies the claim level on each, because “aligned to” and “certified against” are different assurances, and a firm that blurs the two is itself a finding worth noting.
Which Frameworks Apply to a Singapore or Australian AI Deployment?
For a Singapore deployment the binding law is the PDPA where personal data is processed, alongside the voluntary IMDA AI Verify framework; for an Australian deployment the binding law includes the Privacy Act 1988 and the Australian Privacy Principles for covered APP entities (with consumer, discrimination, cybersecurity and sector laws applying where relevant), alongside the non-binding DISR Guidance for AI Adoption (classified “Non-binding” by the OECD.AI Policy Observatory). Both can map to ISO/IEC 42001 and the NIST AI Risk Management Framework as international reference points. The AI Verify Foundation’s testing framework assesses 11 internationally recognised governance principles through documentary evidence and crosswalks to ISO/IEC 42001. Australia’s DISR Guidance for AI Adoption sets six essential practices, published on 21 October 2025 as updated and simplified guidance that, according to DISR, evolves the Voluntary AI Safety Standard and its ten guardrails. Which instruments apply is entity-, activity-, data- and sector-dependent, and is assessed against the client and supplier entities, the affected individuals, the processing locations and data flows, the sector and the deployment context rather than the deployment postcode alone.
Don’t Make These Selection Errors
The costliest selection error runs the meeting before the checks, and the deck-builder wins every meeting it reaches. The correction is procedural: qualify on the seven checks first, and let only the firms that pass through into the room.
Over 40% of agentic-AI projects are forecast to be cancelled by the end of 2027 on escalating costs, unclear value and inadequate risk controls, per Gartner’s 25 June 2025 press release from Sydney, evidence of implementation and scaling risk rather than proof that demo-led selection causes cancellation, though production discipline is what this guide argues reduces the exposure. Every comparison a shortlist needs sits inside the same comparison library this checklist belongs to.
The Brand-As-Shortcut Error
Choosing a firm on brand reputation substitutes a name for the seven checks, and a recognised brand fails the same delivery-evidence test as an unknown one. Brand-led selection, treating a Big Four practice or any other recognised name as a shortcut past the seven checks, is the disqualified method on this page, and no individual firm gets named, ranked or dismissed here. The buyer runs the matrix on the brand-name firm exactly as on the boutique, because the checks are firm-agnostic. How Big 4 practices and boutique agencies actually differ on delivery runs that same matrix against firm class, for a buyer weighing brand against delivery directly.
The Demo-As-Evidence Error
Treating a demo as proof of capability rewards the firm that stages the best demo, not the firm that ships to production, because a demo passes evaluation by accident, while production passes by discipline. The correction is the production record: demand uptime or audit evidence for a system running past month 12, not a live demo staged for the pitch.
The Price-First Error
Comparing quotes before scoping the work compares unscoped numbers as if they were equivalent, and the cheapest quote usually omits the most. The correction is a scoped, itemised quote from every firm on the shortlist, so the comparison runs like-for-like on deliverables rather than a race to the lowest headline figure; what a scoped AI consulting engagement costs at market rates sets the third-party benchmark a buyer checks a quote against.
The Consultant-Vs-Hire Skip
Choosing a consulting firm before deciding whether to engage or hire skips the prior question, and the right answer to “which firm” depends on whether an external partner is the right route at all. The correction is to settle whether to engage a firm or build in-house first, then run the seven-check matrix on the shortlisted firms once the sourcing route is decided.
See How synthesised.ai Meets These Standards
Because delivery evidence is a strong measure of an implementer, a synthesised.ai engagement is structured to answer all seven checks (delivery artefacts, governance evidence and a scoped quote) through the enquiry form. Synthesised.ai engineers production AI with eval-loops, an observability layer and a governance evidence pack that the delivery method is designed to build in each sprint, so, in the illustrative shape of the work, the system that ships early is the same system that stays audit-ready later where a regulated use case applies, with milestones and commitments agreed per engagement. Shipped, not demoed.
Framework relationships are described as building to the relevant controls of ISO/IEC 42001, the NIST AI Risk Management Framework, IMDA AI Verify and DISR guidance; building to a framework is not certification. Written by Kevin Dam, Founder, Synthesised AI Pte. Ltd. Synthesised.ai builds its delivery to the relevant controls of ISO/IEC 42001, the NIST AI Risk Management Framework, IMDA AI Verify and DISR guidance, and publishes no client logos or testimonials without written consent. No pricing figure appears on this page; the scoped quote arrives through the enquiry form once a firm clears the seven-check matrix above, synthesised.ai included.
See how an synthesised.ai engagement answers all seven checks.
Start a scoped conversation
Every commercial step begins with a scoping conversation, and a written proposal follows it. No list prices appear on this site. Tell us the problem, the stack you already run and the timeline.
Book a discovery call →