Buying a data analytics consulting firm is the one enterprise purchase where the review pool is empty. You can read 400 G2 reviews before signing a $40,000 software contract, and almost nothing before signing a $400,000 analytics program. When we pulled Clutch profiles for the six US-facing firms in our data analytics consulting roundup in July 2026, one had two reviews and the rest had zero.
So the buyer does what the vendor prefers: reads a capability deck, watches a case study slide with the client name redacted, checks that somebody called the firm a Leader, signs. Then discovers in month three that the senior data scientist from the pitch is on another account and the model is being built by two people 14 months out of grad school.
This guide is the alternative, for the analytics director, supply chain VP or CDO who has to pick a firm and defend the spend to a CFO who has read the same headlines about AI projects going nowhere.
The decision you are actually buying
Write down the business decision the analytics work has to improve before you write a requirements document. Not “modernize our analytics.” The decision. Which SKUs to build inventory for in September. Which 3% of accounts to hand to retention. Whether the promo lifted anything.
Attach the current number and the error you live with. A forecast running 22% MAPE at the DC level is a starting line you can be measured against. “Better visibility” is not, and it is the phrase that lets a firm bill you for 18 months without ever being wrong.
That sentence also sorts the market. A decision that needs a model is a data science engagement. A decision that is right but arrives four days late is a data engineering engagement. A decision nobody trusts because two teams report different revenue is a governance and semantic-layer problem, and no amount of machine learning fixes it. Word the mandate vaguely and a firm will accept it, then staff it with whatever bench is free.
Analyst recognition, read the way an analyst reads it
Every firm here leads with a Leader badge, and buyers treat the badges as interchangeable. Each program has a different admission rule, a different refresh cycle, and a different thing it measures.
Start with the number that recalibrates everything. Gartner’s Magic Quadrant for Data and Analytics Service Providers, published 15 February 2022, evaluated 18 providers and named 12 of them Leaders, all global SIs or Big Four firms including Accenture, Deloitte, IBM, TCS and Infosys (CX Today’s breakdown , Deloitte’s announcement ). Two thirds of the field were Leaders. That quadrant tells you a firm is large. It does not tell you it is good at your problem.
Everest Group’s PEAK Matrix sorts providers into Leaders, Major Contenders and Aspirants on capability and market impact (Everest Group ). The badge misread most often is Star Performer, awarded on the highest year-over-year improvement between two assessments, which Everest states plainly does not reflect overall market leadership. A firm can be a Star Performer while sitting outside the Leader ring.
The Forrester Wave is invited-only. Forrester picks the vendors, and each participant supplies a questionnaire, a briefing, and reference customers (Wave methodology ). Absence is not evidence of weakness, and presence means the firm cleared a reference check you cannot see, so ask who those references were in your vertical. ISG Provider Lens uses Leaders, Product Challengers, Market Challengers and Contenders plus a Rising Star, and ISG states participation is always free , so the standard suspicion that a badge was bought is answerable program by program.
The badge in the deck is licensed, though. Forrester requires a Reprints license to use the Wave graphic or Leader badge in marketing, plus the statement that Forrester does not endorse any company or product in its research (citation policy ). When a firm shows you a badge, the next sentence is: send me the licensed reprint and the page where my use case appears. Then read the caution paragraph on that firm. It is the only place a paid analyst program writes down what the firm is bad at.
What replaces the reviews you cannot find
Review pools stay thin here structurally. Enterprise analytics work moves through RFPs and multi-year MSAs signed by people whose employer would rather not name the vendor. Software has self-serve buyers. Six-figure consulting does not.
Reference calls you source. Ask for every client in your vertical from the last 18 months, then pick three yourself. Skip satisfaction and ask mechanics: who was named in the SOW, who showed up, how many rotated off in six months, what the first change order was for, what broke in the handover. Then ask what the firm was wrong about. A reference who cannot name one thing was coached.
A paid pilot with a real deliverable. Free proofs of concept get staffed by whoever is between projects and end in a slide instead of a model. Pay for a scoped 60 to 120 day pilot on your own data with an accuracy or lift target written in, and treat the pilot as the evaluation.
A scored RFP that forces method over marketing. Send a redacted or synthetic sample of your real data and ask for the modeling approach, the features they would start from, and the failure modes they expect. Score blind against the weights below. A firm that answers with a capability deck now will answer with one during delivery. We score services the same way we score software, and the method is at /about/methodology/ .
The weighted scorecard, locked before the RFP goes out
Set criteria and weights before any firm presents, get them signed off by finance and security, then let vendors in. A weight you adjust after a pitch is not a weight, it is an excuse. Score 1 to 5, force a written note on any 1 or 5, multiply, total.
The weights sit where services engagements actually fail: hiring a firm with no track record in your problem, and hiring a good firm that sends different people.
| Criterion | Weight | What to score, and the evidence to demand |
|---|---|---|
| Outcome evidence in your use case and vertical | 20 | Case studies with a named client or reference, a stated baseline, and an auditable metric. Redacted logos score 1. Ask for the two engagements closest to your problem, and the one that went badly. |
| Named delivery team, seniority and continuity | 16 | Named leads with CVs in the SOW, senior-to-junior ratio on your account, disclosed attrition, written rotation-notice clause. |
| Capability mix against your real gap | 14 | Engineering, data science and BI are three practices. Make them say which one your program is 70% made of, then check bench depth there. |
| Commercial model and 3-year cost | 14 | Itemized rate card with the onshore/offshore blend, the change-order rate, the run cost after handover. Model any retainer out to 36 months. |
| Independent validation quality | 12 | The licensed reprint, publication year, admission rule, caution paragraph. Platform tiers verified on the platform’s own directory, not the firm’s site. |
| Security, residency and cross-border delivery | 12 | Pass/fail. SOC 2 Type II, ISO 27001, named delivery countries, subcontractor list, DPA with transfer terms, access model for offshore staff. |
| IP, accelerators and exit | 12 | Who owns the models, notebooks and pipelines at the end, what the accelerator licence costs to keep, and the written knowledge-transfer plan. |
Get the Data Analytics Consulting Evaluation Toolkit
The weighted vendor scorecard (Excel, auto-scores your shortlist and ranks the winner) plus the 1-page checklist of questions to ask every vendor and the red flags to walk away from. Free.
Industry depth versus the generalist bench
Vertical depth is the most oversold attribute in analytics consulting, because every firm has a slide claiming all of them. The check: named clients in your vertical from the last 18 months, and how many people on the bench have shipped in it. Across the firms we scored, CPG and retail carry the deepest documented evidence, financial services is patchy, and public healthcare and life sciences evidence is thin nearly everywhere. A strong CPG demand-forecasting record does not validate a payer risk-scoring model.
Domain depth shows up in the questions the firm asks in the first hour. A retail analytics lead who has done this asks about promo cannibalization and returns lag before asking about your cloud. If the first working session is all architecture and no business vocabulary, you are buying a generalist bench with a vertical slide.
Proof of outcomes versus the logo wall
Treat the logo wall as decoration. A logo means someone there signed something once, possibly a two-week assessment five years ago.
An outcome claim earns points only with three parts: the baseline before, the number after, and someone who will confirm it on a call. “Improved forecast accuracy” scores nothing. “Cut DC-level forecast error from 24% to 17% MAPE across 4,000 SKUs, here is the client’s supply chain director” scores five.
Be strict, because the base rate is bad. Gartner predicted at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, on poor data quality, weak risk controls, rising cost and unclear business value (Gartner, July 2024 ). MIT’s NANDA study put it harder: 95% of the GenAI pilots it examined produced no measurable P&L return, across 300 disclosed initiatives, 150 leadership interviews and 350 employee surveys (The GenAI Divide, 2025 ).
Neither failure mode is about model quality. Projects die on data quality, adoption, and nobody owning the decision, and Gartner puts the average annual cost of poor data quality at $12.9 million per organization (Gartner ). Ask every firm what happens in week three when the data is worse than your RFP claimed. The good answer is a named remediation phase and a re-baselined timeline. The bad answer is confidence.
The team you get after signature
You are not buying a product with fixed behavior. You are buying an allocation of people that changes every quarter, and the pitch team is a sales asset.
Get names in the SOW: engagement lead, lead data scientist or architect, and the named senior in your vertical, each with a committed time percentage and a notice period before rotation. Require replacements at equal grade with your written approval. Without that clause, “we have 6,800 people” means two of them, and neither is the one you met.
Ask for last-twelve-month voluntary attrition, then compare it to something real. The listed Indian IT majors that anchor this delivery model disclose it quarterly: Infosys reported IT services LTM voluntary attrition in the mid-teens through FY26 (Q2 FY26 fact sheet ), against a 27.7% peak in March 2022 (Q4 FY22 fact sheet ). A private firm quoting single digits is either unusually good or measuring something else, so ask for the definition and the period. Two firms in our roundup are publicly listed, which puts headcount and attrition in filings you can read without permission.
Then the unglamorous question: how many people on my account are dedicated versus shared, and shared with how many other clients. A senior at 20% allocation is a reviewer, not a builder.
Accelerators, IP, and what you own at the end
Every firm here sells accelerators: migration tooling, forecasting frameworks, industry data models, platform specializations like the Databricks Brickbuilder programs. They are genuinely useful and they are also a lock-in mechanism, so treat the licence terms as part of the price.
Ask four things. Which parts of the deliverable are pre-existing firm IP versus work made for you. What happens to those parts when the engagement ends, perpetual licence, annual fee, or the pipeline stops working. Whether the models, notebooks and pipelines built for you are assigned to you in writing. And whether your team can maintain the result without the firm’s proprietary layer in the middle. An accelerator that saves two months and then charges rent forever is not a discount, so price it over three years against the alternative.
Engagement models and what each one really prices
Time and materials prices hours, so the firm carries no delivery risk and you carry all of it. Fine for exploratory work where scope genuinely cannot be written, dangerous the moment it can. If you sign T&M, cap it monthly and require a burn report against a named deliverable.
Fixed-bid prices a scope, so the firm carries the risk and buffers for it. Right shape for a pilot. The trap is the change-order rate, so get it into the original SOW, because that is the number you negotiate from once the data surprises everybody.
Outcome-based pricing sounds like the answer and usually is not, because the outcome depends on decisions you control. Tie fees to a forecast-accuracy target and you have to agree measurement method, data cut, exclusions and seasonality window up front, in writing. The version that works is a hybrid: a smaller fixed fee plus a bonus on one metric both sides can audit.
Managed analytics, an embedded pod on a monthly retainer, is where most enterprise programs end up. Efficient, and sticky. At roughly $60,000 to $200,000 a month for 10 to 25 people, a two-year managed engagement is a multi-million dollar decision that often gets approved as an extension. Put a 12-month review and an exit-assistance clause in on day one.
The three-year cost, with the lines nobody quotes
The proposal number is a starting position. Build this model before your first CFO conversation, using the firm’s written figures where they exist and marked assumptions where they do not.
| Cost line | Typical range | Where the number comes from |
|---|---|---|
| Discovery or assessment phase | $10,000 to $40,000, 2 to 3 weeks | Scoping tiers published by firms in our 2026 roundup |
| Scoped pilot, 60 to 120 days | $50,000 to $300,000 fixed bid | Published pilot bands, verified July 2026 |
| Phase 1 build, 5 to 8 months | $200,000 to $1.2M | Published Phase 1 bands, verified July 2026 |
| Managed or embedded team | $60,000 to $200,000 per month, 10 to 25 people | Published retainer bands, verified July 2026 |
| Change orders and re-baselining | 15% to 20% of the signed SOW | Topickz planning guidance for this category |
| Platform, tooling and cloud consumption | Not in the proposal. Demand a written monthly estimate. | Warehouse sizing, pipeline frequency and retraining cadence are the partner’s choices and your bill |
| Your side: SME hours, data access, security review | Rarely budgeted, routinely why a pilot slips | Internal, modelled as loaded hours per week |
| Run cost after handover | Median US data scientist wage was $112,590 in May 2024, plus employer overhead | BLS Occupational Outlook Handbook |
Two lines get argued about later. The blended rate is a weighted average of onshore and offshore hours, so a competitive-looking rate can hide a mix that is 80% offshore junior. Ask for the rate card by grade and location plus the assumed mix, then recompute the blend yourself. For an outside anchor on US labor categories, the GSA’s Contract-Awarded Labor Category tool publishes fully burdened hourly rates actually awarded on federal schedules, free.
Ramp is the other. The first three to six weeks are partly the team learning your systems at full rate, so negotiate a discounted ramp or a fixed-fee onboarding block. Price the exit before you sign the entry too: knowledge transfer, documentation standards, repository handover and 30 to 60 days of post-engagement support are cheap as clauses and expensive as change orders.
Security, residency and the cross-border question
For services this is a gate, not a scoring line, and it is broader than the software version because humans in another country get read access to your production data.
Ask for the firm’s own SOC 2 Type II report with its scope stated, a current ISO 27001 certificate, and the delivery model: which countries, which legal entities, which subcontractors. Every firm we scored has substantial India-based delivery behind its US offices. Normal, and also the thing to write into the MSA rather than assume from a New Jersey or Santa Clara address.
Get the access architecture in writing. Virtual desktops with no local download, individual named accounts instead of shared service credentials, background checks to your standard, revocation within a stated number of hours of a roster change. If any answer has data landing on a laptop, that is a finding.
The delivery-side floor also moved. India notified the Digital Personal Data Protection Rules on 13 November 2025, with substantive obligations phasing in and the main compliance date landing 13 May 2027 (Press Information Bureau ). Ask how the firm is preparing, and get transfer terms and breach-notification windows into the DPA now instead of renegotiating in 2027.
The buying committee, mapped
These purchases stall in rooms, not in evaluations. Name everyone who can say no, then pre-write the evidence that answers each of them.
The CFO wants payback and the shape of the multi-year commitment, especially the retainer that quietly renews, so bring the three-year model with the managed-team line extended to 36 months. The CDO or analytics VP wants capability fit and does not want to be embarrassed in month four: bring the scorecard and the references you sourced yourself. Security and privacy want the delivery-country map, the access model and the SOC 2 scope.
Procurement and vendor risk want financial stability, insurance, and the named-team and exit clauses. The business owner whose decision this improves wants to know what changes on their Monday report, and their disinterest is what kills adoption after go-live. Legal wants IP assignment, the DPA and termination terms, so send them the SOW early: ambiguous IP ownership is the most common late-stage reason these deals slip a quarter.
Red flags that should end an evaluation
Some findings are not deductions, they are exits. A logo wall with no client who will take a reference call. A Leader claim with no licensed reprint behind it, or one whose report turns out to be four years old. A proposal with no named delivery lead, or names with no committed time percentage.
A refusal to state the onshore/offshore mix behind the blended rate. An IP clause that leaves the firm owning what you paid to build. “TBD” where the change-order rate should be. And the quiet one: a firm that answers your data-quality question with reassurance instead of a remediation phase.
Any of these is the firm showing you how the relationship works once the invoicing starts.
Questions buyers ask before they sign
How do I evaluate a data analytics consulting firm with no G2 or Clutch reviews?
Use three harder sources: reference calls in your vertical that you pick from a full client list, a paid 60 to 120 day pilot on your own data with a measurable target, and a scored RFP that asks for modeling method against a real data sample. Then check the validation properly, meaning the licensed analyst reprint, its publication year, and the caution paragraph about that firm.
What does Everest Group Leader or Forrester Wave Leader status actually prove?
That the firm cleared that program’s bar at that point in time, nothing more specific. The Forrester Wave is invited-only and requires reference customers, ISG says participation is free, and Everest’s Star Performer badge measures year-over-year improvement rather than leadership. For scale, Gartner’s 2022 Magic Quadrant in this category named 12 Leaders out of 18 firms evaluated.
What should a data analytics consulting engagement cost in year one?
From the bands firms in our roundup published as of July 2026: discovery $10,000 to $40,000, a scoped pilot $50,000 to $300,000, a Phase 1 build $200,000 to $1.2M, embedded teams roughly $60,000 to $200,000 a month for 10 to 25 people. Add 15% to 20% for change orders, plus the platform consumption the partner’s design decisions drive.
How do I stop the senior team from disappearing after the SOW is signed?
Contract it. Name the engagement lead and lead scientist with committed time percentages, require written notice and your approval before rotation, and require replacements at equal grade. Ask for last-twelve-month voluntary attrition and its definition, then ask how many other accounts each named person sits on.
Data engineering firm or data science firm, which one do I need?
Diagnose the failure first. If the decision would be right but the data arrives late, incomplete or inconsistent, you have a pipeline and governance problem and want engineering depth. If the data is trustworthy and the decision itself is weak, you want decision science. Needing both under one SOW is fine, as long as both practices are genuinely staffed rather than one being a slide.
What do I need in the contract about my data leaving the US?
Name the delivery countries and legal entities, list permitted subcontractors with a change-notice requirement, and set the access model: virtual desktop with no local download, individual named accounts, background checks, revocation within a stated window. Add a DPA with transfer terms and a breach-notification window, and keep India’s DPDP Rules on the agenda since substantive obligations land 13 May 2027.