How to Evaluate Data Engineering Services: The Partner Scorecard You Defend to a CFO

You have a data platform problem you cannot staff your way out of. The warehouse migration has slipped two quarters, your two data engineers spend most of the week babysitting pipelines that break at 4 a.m., and the last senior req sat open for five months. So someone above you says: go find a firm.

Then the pitch meetings start, and every deck is the same deck. Partner badges across the top slide, an accelerator diagram in the middle, a logo wall of Fortune 500 clients at the end. Four firms, four nearly identical stories, and no honest way to tell which one puts good engineers on your account and which one puts the B team.

That is the actual evaluation problem with data engineering services. You are not buying software you can trial for fourteen days. You are buying a group of people you have not met, priced by a rate card built to hide who they are.

This guide scores the firm instead of the deck. The weighted scorecard, the engagement economics including the cloud bill nobody quotes you, the staffing clause that stops a bait-and-switch, the security gate for a vendor holding your warehouse credentials, and the one page that gets a signature.

Grab the scorecard and the question checklist near the top and fill them in as you read.

The problem statement you write before the RFP

Write down the failure before you write the RFP. Not “we need a modern data platform.” The specific thing that is broken, in numbers, in one paragraph.

Finance closes the month on a spreadsheet because the revenue model in the warehouse disagrees with the ERP. Six of your eleven pipelines have no owner and no tests. Your analysts wait nine days for a new source system to land. A single pipeline failure last quarter sent the wrong inventory numbers to three regional managers.

That paragraph does more work than a 40-page requirements document. It becomes the acceptance criteria on the SOW, and it is what your CFO measures you against in eighteen months. A soft problem statement produces a soft scope, and a soft scope is how a $300K Phase 1 becomes a $700K Phase 1.

Then decide the shape of the engagement, because firms are built for different ones. A one-time migration with a clean handoff is a fixed-bid build. Ongoing pipeline development you cannot hire for is managed capacity. An architecture you do not trust is a four-week assessment, nothing more.

Those are not the same purchase and not the same vendor. Firms with 3,000 engineers are built for years three and four of a program. Boutiques are built for year one. Decide which you are before a partner decides for you.

The weighted scorecard, locked before the pitch meetings

Here is the mistake almost every services buyer makes. They sit through four pitches, get impressed by the firm with the best partner badges, then build a scorecard that happens to reward partner badges. Set the criteria and the weights first, get the head of data and procurement to sign them, then let firms present.

Score every firm 1 to 5 on each criterion. Force a written sentence behind any 1 and any 5, so “they seemed sharp” cannot hide inside a number. Multiply, total, rank.

The weights below sit where services engagements actually fail. Who is on the team and whether they stay carries the most, because a Snowflake Elite badge on a firm that staffs your build with three engineers eight months out of bootcamp buys you nothing. Certifications matter, and they are the easiest thing on this list to fake with a training budget.

The downloadable version does the math across up to five firms and flags the winner.

Two of those weights get argued every time. Somebody on your committee will want partner tier at 25 and named team at 8, usually the person who has never inherited a codebase from a consultancy. Hold the line. The tier tells you the firm can escalate to Snowflake or Databricks product engineering. It tells you nothing about who writes your dbt models.

The pyramid inside the blended rate

Services pricing has one structural trick and every firm uses it. You get quoted a blended hourly rate, or a fixed-bid number that implies one, and the blend conceals the staffing mix underneath. A $150 blended rate can be one principal architect and one senior engineer, or it can be one architect carrying four juniors. Same rate, wildly different code.

Public rate data shows how wide that spread runs. Accelerance’s 2026 global rates guide , built from more than 100 firms, puts junior developers in Latin America at $33 to $45 an hour and seniors at $60 to $75. In Asia the same guide reports juniors at $24 to $31 and seniors at $31 to $41.

So a firm can staff your pipeline work at the bottom of a band, bill at the top of a US blended rate, and book the difference. That is the business model, not a scandal. Your job is to price the mix rather than the blend.

Ask for it in this exact form: hours at each seniority band, the rate at each band, and a name attached to each senior seat. Firms that intend to deliver with senior people hand that over without blinking. Firms that do not will explain that they price on value, not hours, which is the polite version of no.

Public rate bands do exist for some firms. Clutch publishes them in fixed buckets, and when we pulled provider profiles for our 2026 review of US data engineering firms on June 2, 2026, most specialist consultancies disclosed nothing at all. The two that did sat at $100 to $149 and $150 to $199 an hour.

For a fully burdened US benchmark, the federal government publishes awarded labor rates by labor category in GSA’s CALC+ pricing tool , free to search.

The three-year engagement cost, including the bill you did not sign

The Phase 1 fixed-bid is not the budget. It is the entry fee. Here is a three-year model for a mid-market platform build followed by a managed run, with the source behind each line so you can swap in your own quote.

That last line is the one buyers miss and CFOs find. Every architectural decision your partner makes lands on your Snowflake or Databricks invoice, and they are not the ones paying it. A pipeline that reruns hourly instead of on arrival, a warehouse sized two steps up “for safety,” materialized views nobody prunes. All of it is monthly cost the SOW never mentions.

The base rate of expansion here is public. Snowflake reported a 125 percent net revenue retention rate for fiscal 2026 , meaning the average existing customer spent 25 percent more year over year, with 733 customers past $1 million in trailing twelve-month product revenue.

Databricks crossed a $5.4 billion revenue run-rate in February 2026, growing over 65 percent . Platform spend compounds, and consultancies write the queries that compound it.

Put a consumption forecast in the SOW: month one, month six, month twelve, assumptions written down, actuals reviewed against it in the monthly steering meeting. Firms with a real FinOps practice hand you a cost waterfall by pipeline. Firms without one send a screenshot of the credit dashboard, which has never been the same thing.

The staffing bait-and-switch, and the clause that stops it

This is the most common failure in services buying and it is almost never named in the RFP. The pitch comes from a principal with fifteen years of experience. The SOW says “one lead architect, two senior data engineers, two data engineers.” Week three, the principal is at 5 percent allocation, and the two seniors are one senior plus somebody who finished onboarding last month.

Attrition drives part of it and deserves pricing rather than moralizing. The large publicly reporting IT services firms disclose the number quarterly: Infosys reported 13.0 percent last-twelve-month voluntary attrition in IT services for the June 2026 quarter, TCS 13.6 percent.

Across a five-person pod over eighteen months, that is close to a coin flip on your lead architect still being there at handoff.

The fix is contractual, and firms accept it more often than buyers expect because most buyers never ask. Three clauses:

A named-team clause. Specific human names in the SOW, not role titles, with a written statement that none of them is on another full-time commitment at signing. Substitution requires your written approval and a resume of equal or greater seniority.

A seniority floor with a rate consequence. If a senior seat gets filled by someone below that band, the rate for that seat drops to the lower band for the duration. This one clause removes the entire financial incentive behind the swap.

A ramp-down on ramp-up. The first two weeks of any replacement engineer bill at 50 percent, because you are paying them to learn a codebase your money already built. Every firm will tell you ramp-up is unavoidable. Correct. It is also not something you should pay full rate for twice.

Ask for all three pre-redline. A firm that pushes back hard on the seniority floor has told you what they were planning.

The security gate for a vendor holding your warehouse keys

A data engineering firm is not a normal vendor. They do not process a copy of your data in their own tenancy. They get credentials into the system where all of it lives, usually with write access, often with more privilege than your own analysts hold. Treat this as a gate the firm clears or does not, never a criterion where they lose a point and still win.

The threat model is not theoretical. Verizon’s 2026 DBIR found third-party involvement in 48 percent of breaches, up 60 percent year over year, with vulnerability exploitation at 31 percent now the leading entry point. Your third party has SELECT on every customer table you own.

Demand documents, not assurances. The firm’s own current SOC 2 Type II report with scope and trust services criteria, which is a different artifact from the SOC 2 held by Snowflake or Databricks. A signed DPA before a single credential is issued. Named data residency, written into the MSA rather than the SOW, since MSA terms govern what a SOW can contain.

A written subprocessor list you can object to, which matters more here than in software because subcontracting a pod to a partner firm is routine. Under GDPR Article 28 a processor cannot engage another processor without your authorization, and the same obligations flow down. Background checks confirmed on every named engineer. Cyber liability limits your legal team signs off on.

Then the access model, where these engagements get genuinely sloppy. Individual accounts through your SSO, never a shared service account the pod passes around. Least-privilege roles scoped to the schemas in scope. No production PII in development, verified by you rather than promised by them.

Session logging you can audit. And a credential revocation SLA with a number in it, because the day an engineer rolls off is the day their access should die.

The buying committee for a services purchase

A software purchase has one economic buyer. A services engagement at this size has five people who can stop it, and the deals that stall are the ones where the champion mapped the firm but not the room. Name each of them and the one piece of evidence that answers their objection.

The Head of Data or CDO inherits every architectural decision this firm makes and lives with it after the pod leaves. Bring the architecture decision records, the documentation samples, and the exit plan. This person is your strongest ally or your loudest blocker, usually depending on whether you involved them before the shortlist.

The CFO wants payback and the shape of the multi-year commitment: bring the three-year model with the consumption line in it and the escalation you capped. The CIO cares about standards and who owns the platform in year two. The CISO cares about a third party with warehouse credentials, and the evidence pack is the security section above.

Procurement and legal care about the MSA: named-team clause, residency, rate cap, IP ownership on accelerators, termination for convenience. Get them the draft early. The last stakeholder is the internal engineer who inherits the runbooks, and nobody invites them until the handoff goes badly.

The technical interview that replaces a trial

You cannot trial a consultancy, so build the closest substitute. Two hours with the proposed delivery team, not the sales partner, against a real problem from your environment. That single meeting is more diagnostic than any reference call.

Three parts. Forty minutes of architecture whiteboard on a constraint you actually have, with your own senior engineer asking follow-ups. Forty minutes of code review where they walk you through a real pipeline they built, tests and all. Forty minutes on the ugly stuff: a failed backfill at 2 a.m., their on-call model during your engagement, a source schema that changes without notice.

Watch two things. Whether the lead architect asks better questions than you expected, because good ones surface a constraint your RFP missed. And whether the senior engineers talk at all, or the partner answers for them.

Then call references you picked. Ask for every client in your vertical from the last eighteen months and choose three off that list. One question does most of the work: knowing what you know now, would you sign the same MSA again? The pause before the answer tells you more than the answer.

The one page you take to the CFO

One page. Not the scorecard, not the proposal, not your notes. One page a finance leader reads in ninety seconds and approves.

The recommendation and the number in the first line: the firm, the Phase 1 fixed-bid, the three-year exposure. The problem statement from day one with its number intact. The three-year model laid out by year, so the managed-run curve and the rate escalation are visible rather than buried.

The top risk, which for services is staffing continuity, and the three clauses you negotiated against it. The consumption forecast with its assumptions. One line on why this firm over the runner-up, lifted from the scorecard.

This page works because it argues in the CFO’s units instead of yours. They are pricing a multi-year services commitment against payback and downside exposure, in a year when Gartner expects worldwide IT services spending to pass $1.87 trillion and every function in the building is asking for a partner.

Be blunt about the downside too. Gartner predicts 80 percent of data and analytics governance initiatives will fail by 2027 . Do not project an outcome that ignores that base rate. Bring a scope small enough to prove and a plan for what you own after the pod leaves. That is how phase two gets funded.

Red flags that should end a vendor conversation

Some findings are not point deductions. They are exits. A firm that will not name the delivery engineers in the SOW. A rate card with one blended number and no seniority breakdown. A partner tier claimed on the firm’s website that does not appear in the Snowflake or Databricks directory when you check it live.

A migration line quoted as TBD. No references in your vertical, or references the firm insists on selecting. Refusal to put data residency in the MSA. Documentation delivered as a change order rather than as part of the build. A security questionnaire that comes back in marketing language instead of documents.

Any one of these is the firm showing you how the relationship runs once the money is committed. Believe the preview.

For how we score and verify the firms we write about, see our methodology .

Questions buyers ask before they sign

How do I evaluate a data engineering firm without getting sold by the partner in the pitch?

Lock a weighted scorecard before the first meeting and put the heaviest weight on the named delivery team rather than on partner badges. Then run a two-hour technical interview with the engineers who will actually be on your account, using a real constraint from your environment.

The person who pitches you is a sales asset. The people in that technical interview are the product. Score the product.

What should a data engineering engagement actually cost over three years?

Model the partner side at roughly $1.5 million for a mid-market Phase 1 build plus eighteen months of managed run, using a $400K Phase 1, a 15 percent change-order contingency, $45K a month managed, an uncapped year-three increase, and a knowledge-transfer line. Then add your own half-time internal owner and the cloud consumption the architecture creates.

The two lines buyers leave at zero are consumption and the internal owner. Both are real, and a CFO will find them.

How do I stop a firm from swapping senior engineers for juniors after signing?

Three clauses, negotiated pre-redline. Name the individuals in the SOW with a statement that none is on another full-time commitment. Add a seniority floor where a downgraded seat bills at the lower band. Bill the first two weeks of any replacement at 50 percent.

The middle clause is the one that matters, because it removes the margin that makes the swap profitable. A firm that fights it has told you the plan.

Should I ask for a fixed bid or a time-and-materials rate card?

Fixed bid for anything with a defined deliverable, such as a migration or a first platform build, because it forces the firm to own its own estimate. Time and materials or managed capacity for ongoing pipeline work where the backlog genuinely changes month to month.

Where buyers get hurt is a fixed bid on a scope nobody defined. That is a change-order machine with a reassuring number on the cover.

What security documents should I demand from a data engineering partner?

The firm’s own current SOC 2 Type II report with scope, not the platform vendor’s. A signed DPA before any credential is issued, a named data residency region written into the MSA, a subprocessor list with a right to object, background-check confirmation on every named engineer, session logging you can audit, and a credential revocation SLA for roll-off.

Then the access model itself: individual SSO accounts, least-privilege roles per schema, and no production PII in development.

Does the Snowflake or Databricks partner tier actually mean anything?

It means something specific and narrow. Snowflake sets services tiers of Select, Premier, and Elite based on certifications, pipeline, and customer success, and Databricks runs Registered, Select, Elite, and Global Elite on similar mechanics. A high tier buys faster escalation to product engineering and earlier roadmap visibility, which is genuinely useful on a novel build.

What it does not tell you is who writes your code. Check the tier in the vendor’s own directory on the day you shortlist, since firm websites lag real tier changes, then ask how many of the cited certifications belong to people on your account.

Should I hire in-house data engineers instead of engaging a firm?

Both, in sequence. Firms win on greenfield builds and migrations where you need six people for five months and then need zero. In-house wins on the run, where the work is continuous and institutional memory is the whole value.

The BLS median wage for database architects was $135,980 in May 2024, so two in-house engineers cost less annually than a mid-size managed engagement. The pattern that works is a firm for the build with a contractual knowledge-transfer deliverable, then two or three hires who own the platform from year two.