How to compare AI implementation proposals

Compare scope before selecting a bidder

AI implementation proposals often describe different projects while appearing to compete for the same budget. One includes data cleanup, another assumes it is finished, and a third promises an outcome without defining how it will be measured.

Do not start by ranking presentation quality or dividing the fee by estimated hours. First normalize every proposal into the same evidence sheet. That makes omissions visible without pretending every provider must use the same technical approach.

Issue a common comparison brief

Give each bidder the same process description, system inventory, volume assumptions, constraints and intended business result. Identify known data problems and dependencies on your own staff.

Define the boundary of the purchase: discovery only, a production implementation or a phased engagement with separate approval points. Ask bidders to list exclusions explicitly.

Use the consulting and implementation category to organize your search, but evaluate the actual proposal independently. A directory listing is not evidence that a particular engagement will succeed.

Build the normalized evidence sheet

Create one row for each requirement and these columns:

• Requirement identifier and business importance.
• Provider's proposed deliverable.
• Included, excluded, optional or unclear.
• Assumption and customer dependency.
• Supporting evidence and its date.
• Acceptance method and responsible reviewer.
• Fee treatment and contractual location.

Preserve the original proposal wording beside your interpretation. If you rewrite “integrates with CRM” as “synchronizes all custom fields,” you have created a commitment the bidder may never have made.

Send the sheet back for confirmation. An unanswered cell should remain an uncertainty, not become an assumed inclusion.

Separate claims from evidence

Classify support by what it actually demonstrates. A presentation describes an approach. A demonstration shows behavior under selected conditions. A sample deliverable shows documentation quality. A customer reference describes another organization's experience.

None automatically proves performance in your workflow.

The FTC states that advertising claims must be truthful, non-deceptive and evidence-based. Apply that as a purchasing discipline: ask what supports each material capability or benefit claim. This is not a claim that the FTC mandates your procurement method.

For a claimed improvement, request the baseline, measurement period, sample composition, exclusions and calculation. Mark evidence from another setting as contextual rather than equivalent to evidence from your own process. Do not turn an unsupported percentage into a forecast.

Normalize acceptance measures

Replace vague outcomes such as “accurate AI” with definitions that reviewers can apply consistently. Specify the unit being assessed, what counts as correct, what counts as an exception and how unresolved cases are reported.

NIST's AI RMF Playbook recommends identifying testing procedures and metrics to determine whether a system is fit for purpose and functions as claimed. It also calls for documenting risks that cannot be measured.

Translate that into proposal questions: Which cases will be assessed? Who labels the expected result? Who reviews disagreements? What happens if acceptance is not met?

Agree the method before selecting a bidder; otherwise each proposal can appear strongest under its own definition.

Put prices on a common basis

Compare the same scope, planning period, currency and volume scenario. Separate implementation fees, required subscriptions, usage charges, optional work and ongoing support.

For time-based proposals, identify assumptions, estimate ranges and any spending ceiling. For fixed fees, identify exclusions and the change-request mechanism. Neither structure removes scope risk.

Ask whether training, documentation, migration, evaluation and launch support are included. Keep allowances visibly separate from quoted commitments.

Use a cost workbook for detailed calculations, but bring its comparable totals into the evidence sheet. A lower initial fee may simply omit work another bidder priced.

Score evidence without hiding uncertainty

Set weights before reviewing final bids. Possible criteria include scope coverage, demonstrated technical fit, acceptance clarity, delivery approach, operating support and total cost.

Separate mandatory requirements from preferences. A weighted score should not compensate for an unmet essential requirement.

Alongside each score, record confidence and the unresolved question. If two bids rank closely, investigate the assumption most likely to change the decision instead of adding decimal places to the score.

Convert the selection into commitments

Before signing, reconcile the evidence sheet with the statement of work. Confirm deliverables, dependencies, acceptance procedures, payment milestones, change control and remedies.

If a proposal names Workato, Microsoft Power Automate or n8n, verify the proposed configuration rather than inferring capability from the brand. These are platform examples, not companies known to have submitted a proposal to this directory.

Keep the confirmed comparison sheet as the decision record. Its value is a selection you can explain through documented scope and evidence, not merely a winning score.

References and verification

Official guidance checked 2026-10-23. The evidence sheet is original editorial advice, not a mandated NIST form, a hands-on product review or an endorsement.