Ask a general assistant whether a liability clause is acceptable and you get a competent answer about liability clauses in general. Ask the same question about the same clause in the contract you are negotiating this week, against the position your company took in the last forty deals, and the general assistant has nothing to work with. It will still answer. That is the problem.
Ironclad's announcement on October 8, 2026 makes that distinction the product.
The press release describes Ironclad Agent as a conversational place to search contracts and orchestrate specialised agents across every stage of contracting, grounded in the company's commercial context rather than in general knowledge. The mechanism is a proprietary Contract Knowledge Graph that connects clauses, relationships, obligations and past decisions across a company's contracts and workflows.
The detail worth reading twice is the negative half of the sentence. According to the release, the agents draw on negotiation decisions, counterparty positions and internal benchmarks, not generic AI training data or isolated documents. That is a claim about where the answer comes from, and it is the only kind of claim that matters when the output is going to change a contract.
The vendor framing is specific, but the mechanism is standard. Google Cloud's documentation defines it plainly: grounding is the ability to connect model output to verifiable sources of information. Microsoft's equivalent guidance for Azure describes retrieving content to generate grounded answers from your data, and now steers that work toward a knowledge base service.
Put those together and the question for any agent answering questions about your business is not which model it runs. It is which corpus it reads, and whether you can see the passage it used.
The fourth row is where most pilots quietly fail. An ungrounded agent gives an answer that is impossible to audit, so the reviewer either accepts it on vibes or redoes the work. Both outcomes waste the agent entirely. Requiring a citation before an answer counts converts the agent from a narrator into a lookup with a language layer, which is the same principle behind bounded permissions and human approval.
A grounded agent needs three things that no vendor can supply on your behalf: the precedent has to exist somewhere, it has to be reachable by the agent, and it has to be accurate enough to answer from.
That usually means tidying before buying. Contracts in four systems, quotes in a shared drive, decisions in someone's email: the agent will ground on whatever it can reach, including the stale copy. It also means deciding what the agent may touch, and read-only access to a system removes a class of risk that no confirmation prompt fixes, as covered in read-only versus read-write agent access.
Write the corpus and the question set down before the build, so the agent is judged against a list you agreed to rather than whatever it happened to answer well, which is the discipline in spec-driven development with AI agents.
Take three questions whose answers you already know, and one whose answer does not exist in your records. Ask all four. The first three tell you whether grounding works; the fourth tells you whether the agent invents. An agent that says "I cannot find this in the contracts I can see" on the fourth question is more valuable than one that guesses well, because you can put the first kind in front of a client.
Then decide who owns the correction when the corpus is wrong. Someone has to be accountable for the answer, which is the same role design described in why your AI agents need a human guardrail.
If the question is public and evergreen, grounding buys you very little. A request to explain Incoterms, draft a polite follow-up, or summarise a published regulation does not need your internal corpus, and pointing an agent at one mostly adds latency and cost. The dividing line is simple: if the right answer differs for your business, the agent needs your record. If it does not, general knowledge is fine.
The agent retrieves the material it answers from a corpus you can name, rather than relying on what the model learned during training. Google Cloud documents grounding as connecting model output to verifiable sources of information.
Retrieval is the mechanism; grounding is the result. Retrieving documents matters because it anchors the answer in a source you can open and check, which is why citations are the practical test of whether grounding happened.
Usually yes, with caveats. The agent grounds on whatever it can reach, so stale or duplicate copies become wrong answers. The work is deciding which corpus is authoritative, cleaning it, and choosing what the agent may read.
Ask a question whose answer you know is absent from the corpus. An agent that reports it cannot find a source is behaving correctly. One that answers anyway is filling the gap from general knowledge, and you cannot tell which answers those are without requiring citations.
Maxpertise builds and runs the AI team that operates your business. Roles like the AI back office already read from your records, which makes grounding the difference between an assistant that sounds useful and one that can be trusted with a number. The same is true of the account management and qualification work described in back-office AI agents going mainstream and AI for operational efficiency.
In the managed lane we host and run the team, so the corpus, the access rules and the review cadence are our responsibility rather than yours. Both lanes run on the same two human gates: a person approves the plan before anything is built, and a person reviews the result before it ships. The method is at how it works and the options are on pricing.
The adoption research says most small businesses are still experimenting with isolated tools, which is covered in why 66% use AI and 9% embed it. Grounding is usually the step that moves a business from the first group to the second. If that is where you are, start at contact with the questions you want answered and the records that hold the answers.
[1] Ironclad, Ironclad Launches New AI Agents Grounded in Commercial Context, October 8, 2026. Ironclad Agent, the Contract Knowledge Graph, and the statement that answers draw on negotiation decisions, counterparty positions and internal benchmarks rather than generic training data.
[2] Ironclad, Ironclad Agent product page, accessed October 9, 2026.
[3] Google Cloud, Grounding overview, accessed October 9, 2026. Grounding defined as connecting model output to verifiable sources of information.
[4] Microsoft, Using your data with Azure OpenAI, accessed October 9, 2026. Retrieving content to generate grounded answers from your data.
Maxpertise is an AI-native engineering company. We embed native AI engineers inside your team, live in about 10 days. Please enable JavaScript to view the site, or email contact@maxpertise.net.