What to Ask Vendors Before Signing an Enterprise Web Research Contract

Most enterprise web research contracts get signed on the strength of a demo, a reference call, and a discount at quarter end. That is how buyers end up two quarters later with a pipeline that quietly stopped refreshing, a vendor invoice climbing 7% a year, and no clean way to move the workload elsewhere. The demo answered the wrong questions.

By the time you are sitting across from two or three shortlisted vendors with budget approved, the interesting risks are not whether the platform can crawl a site or answer a question about a PDF. They are contractual, architectural, and operational. What are you actually licensing, who owns what the system produces, and what breaks first when scale doubles?

The checklist below is written for the RFP and legal-review stage. It names the questions worth putting in writing, describes what a defensible answer sounds like, and flags the responses that should end the conversation.

Start With the Data Rights, Not the Feature List

The single most consequential clause in an AI research contract is the one describing what the vendor may do with your inputs, your prompts, and the outputs the system generates. Most buyers skim it. They should not. A Stanford CodeX analysis found only 17% of AI contracts clearly commit to following all applicable laws, compared with 36% for standard SaaS agreements. The gap is not accidental. It reflects how new the category is and how much room vendors are quietly reserving.

Ask for four things in writing. First, an unambiguous statement that your queries, uploaded documents, and generated artifacts are not used to train shared or third-party models. Second, a list of every subprocessor, including any hyperscaler LLM API the vendor routes through, and the contractual data-handling terms attached to each. Third, a data-deletion SLA measured in days, not "commercially reasonable" language. Fourth, an assignment of intellectual property in generated outputs to you, the customer.

Watch for vendors who bury a training-data license inside an acceptable-use policy that they can unilaterally amend. The FTC warned in February 2024 that a business collecting user data under one set of privacy commitments cannot unilaterally revise those commitments to enable AI training, and quietly doing so may be unfair or deceptive under the FTC Act. That warning does not stop it from happening. Contract language does.

Read the SLA the Way a Vendor's CFO Reads It

Uptime percentages on the marketing site are the wrong number. The right number is the cap on service credits and the claim process to get them. Enterprise SaaS contracts routinely cap credits at 10 to 30 percent of the monthly fee for the affected service, according to a review of common cloud contract clauses. That is not compensation for a broken pipeline. It is a token gesture with a formula.

A benchmark of enterprise SaaS SLAs found that service credits cap at the monthly fee in 78 percent of contracts, and exit triggers tied to sustained SLA breach appear in only 31 percent. If your workload is time-sensitive — earnings-day extractions, pricing updates, regulatory filings the moment they post — a 25% credit on last month's invoice does not make you whole. Push for three things: named latency and error-rate commitments beyond raw availability, an escalation ladder that reaches an unblocked engineer inside a defined window, and a termination-for-cause right on repeated or material breach.

Two operational questions belong in the RFP as well. What percentage of a customer's crawl targets does the vendor commit to keeping unblocked over a rolling 30 days, and how is that measured? And what is the mean time to recovery when a major target changes its anti-bot posture? Vendors that cannot answer either question are quoting theatre, not an SLA.

Force the Compliance Evidence Onto the Table

Compliance certificates are the closest thing to a shared vocabulary between security teams and vendors. Ask for the artifacts up front, before legal review, because the artifacts themselves take weeks to produce if they do not already exist. Compliance is not optional for most buyers now: 77% of organizations cite standards such as ISO 27001, NIST, and SOC 2 as their top requirement for third-party vendors.

The right ask is a current SOC 2 Type II report under NDA, covering an audit period ending within the last twelve months, with the trust services criteria that match your use case. A vendor that refuses to share a SOC 2 report even under NDA is a significant warning sign, and one that has come up repeatedly in third-party risk reviews. "Working toward certification" is not a substitute for holding one, and neither is a Type I attesting to control design at a single point in time.

Read the exceptions section. An unqualified opinion does not mean zero findings; it means the auditor did not consider the findings material. Your risk team can decide whether they agree.

Get the Scraping and Proxy Policy in Writing

Web research vendors live inside a legal grey zone that has been shifting for the better part of a decade. In hiQ Labs v. LinkedIn, a November 2022 district court ruling found hiQ had breached LinkedIn's User Agreement, producing a $500,000 judgment and a permanent injunction. hiQ Labs subsequently closed. That was the same case in which the Ninth Circuit had earlier ruled that scraping publicly available data does not violate the Computer Fraud and Abuse Act. The lesson for procurement: the CFAA is not the only exposure. Contract law, copyright, and the CCPA all sit on top.

Ask each vendor how they classify targets by legal risk, whether they honor robots.txt, and whether they treat login-gated content differently from public pages. Ask which jurisdictions their proxies operate in and how residential IP consent is obtained upstream. Ask who indemnifies whom if a target files suit. A vendor that shrugs at any of these questions is transferring the risk onto your balance sheet without pricing it in. The reasoning behind a defensible extraction strategy is covered in more depth in this piece on enterprise proxy infrastructure.

Test Reproducibility, Not Just Accuracy

Accuracy demos are staged. Reproducibility is not. A multi-agent research run that produces a different answer to the same question on Tuesday than it did on Monday is not a research system; it is a slot machine with a chat interface. This matters more than it sounds. Production LLM hallucination rates vary significantly by task type, with multi-step agent workflows hallucinating on 20 to 40 percent of tool-call chains. If the vendor cannot show you the tool calls, the retrieved documents, and the prompt tree behind an answer, you cannot debug the 30% that were wrong.

Three concrete asks. Require every completed research task to write a versioned, inspectable artifact — ideally a commit to a repository you control, not a row in the vendor's private database. Require the run to record the exact model version, prompt, retrieval query, and source URLs used. Require the ability to replay a task against a pinned snapshot of its inputs. This is the operational meaning of reproducible research infrastructure, and it is the axis on which serious BYOK vendor comparison is fought.

A useful diligence exercise: hand two candidate vendors the same twenty questions, run them a week apart, and diff the outputs. Then diff the audit trails. The vendor whose trail actually explains the variance is the one to shortlist.

Model the Exit Before You Sign the Entry

The best time to negotiate an exit is before the ink dries. Enterprise SaaS contracts routinely include annual price escalators of 3 to 7 percent, meaning a $50,000-a-year contract at 5% escalation becomes $63,800 after five years — before any usage growth. If the vendor also holds your embeddings, your fine-tuned adapters, and the only working copy of your extraction configurations, the escalator is not really negotiable at renewal.

Bake three provisions into the master agreement. A data-portability clause that specifies export formats, cadence, and a maximum time to fulfill an export request. A source-of-truth clause that names which artifacts belong to you: raw extractions, embeddings, indexes, prompt libraries, agent configurations, and the commit history behind them. And a transition-assistance clause that survives termination for a defined period at a fixed hourly rate. A bring-your-own-key model makes the first two easier by design, because the model weights and the vector store live on infrastructure you already control.

The wider context matters too. RAND Corporation analysis found AI projects fail at a rate exceeding 80%, roughly twice the failure rate of conventional IT projects. That is not an argument against buying; it is an argument for buying in a way that lets you replace the vendor without replacing the pipeline. The economics of doing that are covered in the site's piece on multi-provider routing costs, and the shape of a full 90-day pilot budget is in the POC cost breakdown.

Red Flags That Should Kill a Deal

Some answers are disqualifying, no matter how strong the demo. A vendor that will not name its subprocessors, or reserves the right to change them without notice. A vendor whose "unlimited" credits sit behind rate limits or fair-use clauses that reset the meaning of the word. A vendor whose training-data language grants a perpetual, sublicensable license to your inputs. A vendor whose SOC 2 report is more than eighteen months stale with no bridge letter. A vendor that cannot produce an artifact log for a research run completed in front of you.

Each of those is a signal that the operating model behind the product does not match the contract you are being asked to sign. That mismatch tends to surface first at the moment you most need the vendor to perform: an urgent extraction, a legal hold, a compliance audit, a renewal negotiation. Better to find it in diligence than in production.

What a Good Shortlist Answer Sounds Like

A vendor worth signing tends to answer procurement questions the way an infrastructure team answers postmortems: with specifics, with numbers, and without adjectives. They will hand over the SOC 2 Type II before you finish asking. They will point at their status page and their public incident history rather than a 99.9 marketing figure. They will describe subprocessors, data residency, and deletion timelines in the same breath as pricing. They will show you the commit history of a real research artifact.

If two vendors clear every question above, the tie-breaker is architectural fit, not price. The one whose data model matches how your team already ingests, indexes, and interrogates content — proxies, extraction, embeddings, RAG, agents, and the glue between them — will be cheaper to operate over three years than the one that requires a re-platform. If it helps to see how that stack is assembled end to end, the walkthrough of a next-gen research stack covers the moving parts, and the feature inventory and technical docs describe how those pieces are exposed at the API layer.

Sign the contract that reads well on a Wednesday morning six months in, when the interesting question is not whether the platform is impressive but whether the pipeline is still running and the numbers still tie out.