A workflow that returns a confident wrong answer is harder to detect than one that fails with an error. That is the commercial case for hallucination mitigation: the cost is not the model call, it is the decision taken on an answer nobody verified.
Buying mitigation is therefore not the same as buying a model. You are buying retrieval quality, evaluation, citation rules, refusal behaviour, logging and a human review path around the model. This guide shows how to specify that system and how to tell a credible proposal from a demonstration.
Begin with the failure you cannot tolerate
Write down the worst acceptable outcome for each workflow. A wrong product description wastes a customer's time; a wrong payment instruction, medical claim or compliance statement creates liability. Rank workflows by consequence, not by how impressive the demo looks. A support assistant that drafts replies for a human sits in a different risk class from an agent that can issue refunds.
For each workflow, name the source of truth. If the answer cannot be traced to an approved document, database record or policy, the workflow is not ready for automation. It may still be ready for drafting with review.
Requirements to write before shortlisting
- Grounding sources: which documents, tables or systems the answer must use, who owns them and how often they change.
- Traceability: whether every claim needs a citation, a record link or a page reference, and how the user sees it.
- Refusal rules: the questions the system must decline or escalate, and the exact wording the user sees.
- Evaluation set: real questions with approved answers, including ambiguous and adversarial cases.
- Human review: who checks output, at what volume and with what authority to correct it.
- Change control: notice of model, prompt or retrieval changes that could alter behaviour.
- Logging: what is stored, for how long, who can read it and how it is deleted.
What different vendors actually sell
Model providers sell capability and API access; they rarely own your retrieval or review process. Retrieval and search vendors sell the layer that finds the right passage, which is often where hallucinations begin. Agent platforms sell orchestration, tool access and guardrails. Systems integrators sell the design and the messy work of connecting content, permissions and workflow. Evaluation vendors sell test harnesses and scoring.
Most mid-sized buyers need a combination. The important question is which layer the vendor will be accountable for when an answer is wrong. A proposal that promises "accurate answers" without naming the retrieval source, the evaluation method and the escalation path is selling a demo, not a control.
A scorecard for groundedness
Weight the criteria before demos and require evidence for every score.
- Retrieval precision: can the vendor show the passages used for a sample answer and explain why others were excluded?
- Citation discipline: are citations checked against the cited text, or generated separately?
- Abstention: does the system say "not found" when the source is missing, or does it improvise?
- Evaluation: is there a repeatable test set, a scoring method and a record of failures?
- Observability: can you inspect a single answer end to end, including the model version and retrieved chunks?
- Permission awareness: does retrieval respect the user's access rights rather than the service account's?
- Change management: how are model updates, prompt edits and new documents tested and approved?
Ask for a live failure review. Give the vendor three questions your team knows are hard, and watch how they diagnose a wrong answer. A vendor who can trace the failure to retrieval, prompt or source quality is more valuable than one with a perfect canned demo.
Pricing questions that reveal the cost model
Hallucination mitigation is priced in several ways: per seat, per conversation, per document indexed, per token, per workflow, or as a fixed implementation plus retainer. Ask what happens when document volume doubles, when a second language is added, when the evaluation set grows, and when a model provider changes its price. Ask whether re-indexing, evaluation runs and human review tooling are included or billed separately.
The cheapest per-answer price is irrelevant if the workflow needs a reviewer for every output. Compare the cost per accepted answer, including review time, and the cost of a wrong answer that reaches a customer.
Design the pilot around a challenge set
Build the pilot from real questions, not marketing prompts. Include questions with a clear answer in the source, questions with no answer, questions where two documents conflict, and questions that require a permission check. Run the same set before and after configuration and keep the failure log.
- Agree the decision the pilot supports and the person accountable for it.
- Freeze a source collection and record its version.
- Score answers for correctness, citation accuracy, completeness and safe refusal.
- Measure reviewer time per accepted answer.
- Test a source update and a permission change during the pilot.
- Decide with the business owner whether the residual error rate is acceptable for the workflow.
Contract terms that keep you in control
Require notice before a model or retrieval change that affects production, the right to run your own evaluation after a change, and the ability to export your documents, prompts and logs. Clarify whether your content is used to train shared models, where it is stored and how deletion works. If the vendor cannot explain the sub-processors involved, the data protection review is not finished.
Red flags in proposals
Watch for accuracy claims without a test method, a single aggregate score with no failure examples, "human in the loop" with no defined reviewer or queue, unlimited document ingestion with no ownership of stale content, and a pilot that cannot be repeated without the vendor present. Treat a promise that a newer model will remove the problem as a deferral, not a plan.
Matching the approach to the risk
For low-consequence drafting, a retrieval layer, citations and spot checks may be enough. For customer-facing statements about money, eligibility or safety, require abstention, mandatory review and a full audit trail. For regulated advice, keep a qualified human as the decision-maker and treat the system as research assistance. The technology can be similar; the controls should not be.
How we approach this work
We design and operate AI workflows inside our Business Automation division, including retrieval, review queues and evaluation. The AI-assisted support service covers the customer-facing pattern. Scope and commercial terms are set out on the pricing page. For the vendor comparison that follows this decision, read how AI automation vendor evaluation changes as you grow.
Sources
- NIST, AI Risk Management Framework, accessed 17 November 2026.
- NIST, Generative AI Profile (AI 600-1), accessed 17 November 2026.
- OWASP, Top 10 for Large Language Model Applications, accessed 17 November 2026.
- ISO/IEC 42001, AI management systems, accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.