esedark
Menu
Back to notes

AI consulting / budget / deliverables

AI consulting costs: deliverables and budget components

An assessment, a pilot and a production system buy different things. Ask for a separate budget for each stage and concrete evidence you can use to accept it.

AI consulting costs depend on which uncertainty the engagement must resolve and how much delivery it includes afterwards. An opportunity report is different from a test using your data. A demo does not necessarily include integration, human review or maintenance. I would compare proposals for the same process and obligations before comparing their totals.

This guide helps you budget for an AI assessment before hiring. If you already have a prototype and need a release decision, use the AI pilot evidence review. The question here is what you are buying, which costs remain outside the quote and what closes each budget item.

Decide whether you need assessment, a pilot or implementation

An assessment identifies a task, checks available information and compares alternatives. Its deliverable is a reasoned decision, including reasons to avoid AI. I would not demand proof of accuracy before testing, but I would expect explicit assumptions and a proposed test to resolve them.

A pilot buys evidence about a specific uncertainty: whether agreed fields can be extracted from a defined sample, for example. It should include permitted data, expected answers reviewed by the business, tests held apart from tuning and a report of errors and limitations. A negative result can still complete the pilot if it answers the agreed question. It should not automatically trigger another paid phase.

Implementation connects an accepted capability to real work: permissions, states, review, recovery and operational handover. The guide to integrating AI into a SaaS covers those technical boundaries. Ask the quote to distinguish a disposable experiment from components that will be delivered for production.

Require alternatives before approving spending

I would ask the supplier to compare the current process, rules or integration without AI, an existing product feature and a custom component. Each option should address the same requirement. If a verified configuration solves the task, commissioning configuration and validation may be enough. If the only uncertainty is a connection, scope that check before buying a broad assessment.

The agents versus workflows guide separates text interpretation from autonomy. I would not pay for an agent when the route is already defined. Nor would I start a pilot without an owner to review examples, permitted access to data or a decision its results could change.

Request a non-AI alternative covering the same workflow and exceptions. The business automation service matrix helps assign responsibility for operations, retries and maintenance. Compare the manual work retained by each option, including output review, before approving a pilot: a more complex proposal should explain which decision it improves.

A matrix for comparing costs and accepting deliverables

Copy these rows into your request for a proposal. For each, ask for an amount or calculation method, an owner, included/optional/excluded status, dependencies and completion evidence. An unresolved item is not a zero cost. This is a proposed buying framework, not a supplier ranking or test report.

What each AI assessment budget item should buy
Budget itemDeliverableAcceptance condition
AssessmentWorkflow, current alternative, available data and pilot question.The business validates the workflow and can choose to test, resolve a dependency or stop.
Data preparationAgreed sample, selection rules, reviewed labels and permitted uses.Coverage, gaps and the person confirming expected answers are explicit.
PilotVersioned configuration, results by case type, errors and observed consumption.The question is answered on held-out cases; failures and human review remain in the report.
IntegrationDestination workflow, permissions, exceptions and manual fallback.The owner checks the final outcome and agreed failure and recovery cases.
OperationsRecurring budget, alert owner and change procedure.Usage, human review, support and repeat evaluations are separate; spending overruns have an approver.
Handover and exitAgreed materials, account inventory, instructions and open issues.Another authorised person can run the checks and continue without a supplier's personal account.

Accept a stage when its agreed evidence has been delivered. Authorising the next stage should be a separate decision. For time-based billing, also agree a spending cap and review point: using the allocated hours does not itself establish that the question has been answered.

Example: two quotes for classifying internal requests

Hypothetical example: a business wants to classify internal requests and prepare a draft for a person to review. Proposal A includes interviews and a demo using selected examples. Proposal B also includes preparing expected answers, holding cases apart from tuning and recording corrections. They are different purchases even if both are labelled an “AI pilot”.

To compare them, I would ask A to price those additional items and B to identify its exclusions. If neither connects the result to the team's working tool, both should show implementation as a later decision. I would exclude automatic sending, additional languages and cleanup of the entire historical dataset from this first engagement.

Pilot acceptance would require traceable results on the agreed sample, separated by category, with review effort recorded. The business agrees the threshold for advancing before testing. If one category falls short, it can stay outside the next scope without presenting the remainder as a universal solution. This example contains no client outcomes or measured savings.

Compare total costs without inventing a rate

I would separate the initial work into assessment, data preparation, pilot, integration, agreed migration, training and launch. Effort varies with source count, data quality, permissions, categories, languages, exceptions and who prepares reference answers. I do not quote my own price without understanding that scope.

For the same comparison period, add initial delivery, licences, infrastructure, service usage, human review, maintenance and exit. Avoid counting a licence twice if another item already includes it. Ask for ordinary and peak workload scenarios with explicit assumptions. Where information is missing, commission a bounded measurement and leave the amount unresolved.

Maintenance should distinguish failures against accepted requirements, adaptation to external changes and new capabilities. Ask whether it includes rerunning evaluations when models, instructions or sources change, who supplies test cases and what support coverage you are buying. A response time is not a resolution commitment. The general software quote guide covers milestones and scope changes.

Define exclusions, ownership and responsibilities

List systems, sources, users, environments and permitted actions. State whether full historical migration, ongoing support, additional languages, legal review and certifications are excluded. A technical report should not be sold as a guarantee of compliance or financial outcomes.

The business supplies permitted examples and an owner who validates the criteria. The implementer documents configuration, tests and limitations. Operations handles exceptions. Agree which code, instructions, evaluation sets and editable documents will be delivered, what rights accompany them and which third-party components retain their own terms. Do not assume a demo includes all these materials.

What to send me to scope the engagement

Send one task, your current tools, anonymised input and expected-output examples, approximate volume and the errors that matter. Add who would review the pilot, which access is possible and any budget, timing or maintenance constraints. If you have proposals, their line items and exclusions are enough; leave out credentials and personal data.

I can review the scope of an initial AI assessment and implementation engagement with you. You can also share the process context. The next step is to agree which decision the assessment must resolve and what evidence would justify funding a pilot.

Review my AI budget