Much of the productivity case for AI quietly assumes something no organisation has agreed to.

I use AI to prepare and review the SOWs and contracts passing between ourselves and our customers. Clearly our customers are doing the same. Two organisations, two capable systems, and a process that still takes many weeks.

There is no doubt that the resulting final documents are better because of AI. Inconsistencies are identified and resolved, cross-references corrected, grammar is almost faultless. There is also the meaningful content contained in the documents, such as the liabilities, the commitments, the responsibilities, the commercials. AI helps ensure that these are clearly defined, consistent, complete and add up correctly. Potentially more than may have been the case previously.

So better, yes, and easier in several ways, and possibly faster, although not what you might consider quick. There are certainly new things being introduced that we haven’t seen before or as nearly as regularly. The bias that AI has towards whoever it is working with, whether that is more customer or supplier centric, is clear. It argues our corner when we ask it to, and theirs when they do, and it does both with the same fluency and the same air of reasonableness. That is worth pausing on, because it means what meets at the boundary is not a system and another system. It is two advocates, and no judge.

However, as the latest updated version goes back to the customer or supplier, we meet the 'friction' that AI is supposed to be on the point of removing, by replacing the humans on either side of this and many other similar sorts of trade.

Boundaries 

Inside an organisation, AI is doing real work. Drafting, summarising, translating, restructuring, research that used to take an afternoon. The gains are genuine. Especially to the individual.

But the processes worth the most money nearly all cross a boundary, from person to person, department to department, or organisation to organisation. Internally it is not difficult to see how things can evolve for the better as however it may seem at times, everyone is fundamentally on the same side. This is where the ability to control lies and where the re-engineering of organisations and processes to benefit from AI can most easily take place. It is at the external boundaries, where vested interests diverge, in sales, procurement or contracting for example, that the pause button is hit. The outputs must be read, considered, approved and negotiated by a human, or several humans, who may have a different but not entirely opposing point of view.

Contracts

A first version is sent, with the sender’s bias, which the receiver’s AI reviews, producing a redline and summary. Humans review, update, and return it. The sender’s AI reviews the returned redline and produces a counter. Humans review, update and return. The cycle repeats until agreement or impasse is reached.

Remove the humans and that loop closes in minutes rather than weeks. The systems on both sides are already capable of it. Nothing technical stands in the way.

It is two advocates, and no judge.

So why does nobody do it?

Negotiation

A contract negotiation is not an information-processing problem. You could brief the AI with your positions. But the reason you do not immediately concede an indemnity cap is rarely that you lack a fact. There are countless factors associated with risk, value, cost, insurance, amongst many. Factors that are highly nuanced and inter-related. The agent is optimising the document, but all of this wider context and related ‘rules’ are not easily documented.

There is another problem. A negotiation is largely about not revealing what your final position may be, or what you are prepared to move on in return for something else. Quid pro quo. An agent that responds consistently and quickly is an agent that can be probed with hundreds of variations, mapping your constraint surface. Speed is not neutral here. It actively erodes the asymmetry that negotiation depends on.

So even if you could connect the two systems, you would be automating the part that needs judgement and accelerating the part where slowness was protecting you.

Precedent

This has happened before, several times, and it never resolved through negotiating machines. EDI in retail and automotive. SWIFT in banking. FIX in trading. Open Banking more recently. In every case the machine-to-machine layer worked precisely where the exchange could be reduced to structured messages with agreed semantics. In other words, standards. Everything requiring judgement stayed with people.

Speed is not neutral here. It actively erodes the asymmetry that negotiation depends on.

Purchase orders went machine-to-machine in the 1970s. Contract negotiation still has not. That is not a failure of the technology. It is that a purchase order is structured data and a contract is not.

Authority

Then there is the question of the law, which appears unresolved. If your agent appears to accept a term, have you accepted it?

The law does not necessarily require a ceremonial signature. Apparent authority turns on what the counterparty reasonably believed, given how you presented the thing they were dealing with: your credentials, your systems, your history of honouring what came out of them. An internal instruction capping what the AI may agree to is not obviously a defence if you equipped it to look authorised. And it compounds: every automated transaction you honour becomes evidence of authority for the next one.

Purchase orders went machine-to-machine in the 1970s. Contract negotiation still has not.

In the case Moffatt v Air Canada, the airline’s chatbot told a customer something about bereavement fares that was not true, and Air Canada argued the chatbot was effectively responsible for its own statements. The tribunal disagreed. It is a small claims decision in Canada, not a precedent anyone here is bound by, but it settles the intuition cleanly: “the AI said it, not us” is not a position.

Regulation may eventually settle this properly. It will not settle it soon, and it is unlikely to keep pace.

Connect two agents and let them agree things, and you have manufactured exactly the conditions that argument depends on.

Trust

Confidentiality questions are no easier. Most NDAs permit disclosure to representatives on a need-to-know basis, but is a model a representative? Does putting a draft into a vendor’s system constitute disclosure to that vendor? For contract work specifically, there is a privilege question that will worry your legal team more than the data protection one.

And the shape does not scale. Two organisations is one connection. Ten is forty-five. Which is why we ended up with email standards rather than a private mail link between every pair of companies, and why the agent protocols now converging (A2A, MCP and the rest) are the obvious answer.

“The AI said it, not us” is not a position.

Except that they standardise transport. They do not establish authority, provenance, or who is liable when two agents agree something wrong. The hard problem was never connection.

There is also a risk nobody had before. If my agent reads your draft, your draft is an input to my model, and instructions can be hidden in it. Prompt injection stops being a security concern and becomes a commercial attack surface, in a document you were contractually obliged to accept.

Structure

None of which means agent-to-agent will not happen. It means the automatable part is the structured part, and the work is therefore not building negotiating agents. It is shrinking how much of the process is prose.

Standard clause libraries. Positions expressed as data rather than paragraphs. Structured deviation reporting against an agreed baseline, so that what gets escalated to a human is the handful of terms that genuinely need judgement rather than the whole document.

The organisations that will benefit are the ones whose commercial documents are already structured. Which is the same conclusion I reached writing about digital experience architecture, arrived at from a completely different direction.

Inside your organisation, what AI can do is limited by whether your content exists as structured data or as formatted pages. Between organisations, it is limited by whether the exchange is structured or is prose. Same constraint, different boundary, and in both cases the capability arrived years before the thing that would make it usable.

Until somebody builds that, my AI will read their contract, their AI will read mine, and people will still carry it between them. The people are not the bottleneck. They are the representatives.

 Codehouse designs and builds digital experiences for the AI era.

https://www.codehousegroup.com/