Internal Negotiation
The first serious negotiation between agents may not be between companies.
It may be between two departments that already struggle to agree who owns the work.
That matters because agentic AI is often imagined at the edges of the organisation. A customer’s agent speaks to a company’s agent. A supplier’s agent speaks to a buyer’s agent. A citizen’s agent speaks to a council’s agent. Something is requested, priced, checked, booked, approved or refused.
That may come.
But before our agents negotiate with the outside world, they are likely to negotiate inside our own organisations.
And that may be the more interesting problem.
Internal negotiation is what happens when one agent tries to get something done and another agent represents a different part of the organisation: finance, legal, HR, procurement, operations, customer services, IT or a service area with its own priorities and constraints.
One agent is trying to move quickly.
One is trying to reduce risk.
One is trying to stay within budget.
One is trying to protect compliance.
One is trying to improve resident experience.
One is trying to avoid adding pressure to an already overloaded team.
None of those aims is wrong.
The problem is that they are not the same aim.
In a council, the agent for speed, the agent for compliance, the agent for resident experience and the agent for budget control will not naturally want the same thing. Nor should they. Good organisations contain productive tension. They are not meant to collapse every value into a single instruction.
The danger is pretending that an agentic system can make that tension disappear.
It cannot.
Agents will not remove organisational ambiguity. They may simply process it faster.
That is the uncomfortable point. We may imagine agents as a way to reduce friction, but some friction exists for a reason. Procurement asks questions because public money needs care. Legal asks questions because authority has boundaries. Finance asks questions because affordability matters. HR asks questions because change lands on people. Operations asks questions because someone has to make the thing work on a wet Tuesday afternoon when the project team has moved on.
If those tensions are irritating, that does not make them unnecessary.
An agent that removes them too neatly may not be making the organisation smarter. It may simply be making unresolved conflict less visible.
This is where internal negotiation becomes important.
Imagine a project agent preparing a proposal. It has read the business case, the RAID log, the benefits plan and the latest highlight report. It wants to produce a clean recommendation for the board.
The finance agent pushes back. The cost profile is still too vague.
The procurement agent pushes back. The proposed route may not be compliant.
The legal agent pushes back. The decision wording creates a risk.
The change agent pushes back. The service impact has been described, but not interpreted.
The operations agent pushes back. The implementation date lands in the same month as three other changes.
This sounds useful.
It could be useful.
But it is only useful if the disagreement is visible.
If the project agent smooths the disagreement into a bland paragraph, the board may see confidence where there was actually conflict. If the finance agent blocks the proposal without explanation, the project team may experience governance as obstruction. If the legal agent insists on caution without showing the risk, everyone else may treat legal advice as a brake rather than a guide.
The quality of internal negotiation depends on the quality of the disagreement.
That is true for humans. It will also be true for agents.
The agents may be technically aligned to their instructions, but the instructions themselves may point in different directions. “Move quickly.” “Avoid risk.” “Reduce cost.” “Improve resident experience.” “Comply with policy.” “Protect capacity.” “Do not create work for the service.” “Get this through the board.”
These are not just prompts. They are organisational values colliding.
If we do not name the collision, the system will resolve it somewhere else.
Perhaps it will average the priorities into a weak compromise. Perhaps it will privilege the loudest instruction. Perhaps it will follow the most recent message. Perhaps it will optimise for whatever can be measured. Perhaps it will escalate everything, creating a new layer of machine-generated noise. Or perhaps it will produce a confident recommendation with the conflict hidden inside it.
That is the version I worry about most.
Not the agent that says, “I cannot reconcile these constraints.”
The agent that says, “Here is the recommended way forward,” while burying the unresolved trade-off in polished prose.
That would not be a new failure. It would be a familiar failure with better formatting.
Organisations already do this. We know how to take disagreement, soften it, summarise it and move it into language that allows a meeting to continue. We know how to call a serious dependency “being monitored”. We know how to call a capacity problem “resource constraints”. We know how to call a lack of agreement “further engagement required”.
Agents may learn those habits because we reward those habits.
This is where internal negotiation links to governance psychological safety. If humans are afraid to surface conflict, their agents may learn to smooth it over, hide it in options language or escalate it too late.
An agent will not make an organisation braver.
It may make the organisation’s existing courage, or lack of it, more visible.
That is why internal negotiation needs governance before it needs cleverness.
The question is not simply whether agents can negotiate. They probably can, at least in narrow ways. The question is whether the organisation has decided what good negotiation looks like.
Does good negotiation mean speed?
Does it mean consensus?
Does it mean compliance?
Does it mean surfacing trade-offs?
Does it mean protecting the resident?
Does it mean protecting the budget?
Does it mean making disagreement visible enough for humans to decide?
That last one matters most.
Agents should not be used to settle conflicts that require human authority, public accountability or ethical judgement. They can prepare the ground. They can show the options. They can identify contradictions. They can test whether the evidence supports the recommendation. They can ask whether a decision is still safe.
But they should not quietly decide which organisational value wins.
That belongs to governance.
This is where RACI for machines becomes practical. If one agent proposes, another blocks, a third escalates and a fourth updates the record, RACI+E becomes less of a diagram and more of a survival tool.
Who owns the proposal?
Who owns the challenge?
Who owns the escalation?
Who owns the final judgement?
Who is accountable if the agent-to-agent negotiation hides the very conflict it was meant to reveal?
Without that clarity, internal negotiation becomes fog.
Work appears to move. Messages pass between systems. Exceptions are logged. Recommendations emerge. The machine has been busy. But the organisation may still not know who decided, who challenged, who conceded or who had the authority to resolve the trade-off.
That is not automation.
That is ambiguity at speed.
To make internal negotiation useful, we need to design for visible disagreement.
An agent should be able to say:
“This recommendation is blocked because the procurement route is not yet supported by the evidence.”
Or:
“This option is deliverable only if the service accepts a capacity impact in June.”
Or:
“Finance, legal and operations constraints point in different directions. Human decision required.”
That last sentence is not failure.
It may be the most valuable output in the whole process.
Because the job of the agent is not always to complete the work. Sometimes it is to make the unresolved decision impossible to miss.
This is where agentic AI could genuinely help governance. Not by replacing judgement, but by making the conditions for judgement clearer.
It could show when the same project is being described differently by different teams. It could expose when a delivery date depends on an assumption nobody owns. It could identify when a benefit relies on behaviour change that is not in the plan. It could warn that an implementation date collides with other changes landing on the same service.
In other words, it could help the organisation see itself.
But only if we let it report tension rather than reward it for smoothing tension away.
That is the design choice.
Do we want agents that make organisational life look neater?
Or do we want agents that help us see where the neatness is false?
Internal negotiation will test that choice.
A good internal agent should not be a digital office politician. It should not learn to flatter the sponsor, avoid the difficult service, bury the legal caveat or reword the risk until nobody feels accused. It should not become a machine for producing acceptable ambiguity.
It should be able to hold a line.
Not because it is accountable. It is not.
But because the humans who designed the process decided that some lines matter.
Show the evidence.
Name the assumption.
Expose the trade-off.
Escalate the unresolved conflict.
Record who made the judgement.
Those are not technical niceties. They are the conditions under which internal negotiation remains governable.
The irony is that agents may teach us something about ourselves. If we struggle to define how our agents should resolve internal disagreement, it may be because we have not defined how we expect humans to resolve it either.
Who wins when speed and compliance conflict?
Who decides when resident experience and budget point in different directions?
Who has authority to accept capacity risk?
Who can say that a project should pause because the organisation cannot absorb the change?
These are not AI questions.
They are governance questions that AI will make harder to avoid.
That may be the real value of internal negotiation. It will not simply automate the work. It will expose the hidden agreements and missing agreements that the work already depends on.
If we design it badly, agentic AI will give us faster compromise, smoother ambiguity and more polished uncertainty.
If we design it well, it may give us something much rarer.
Clearer disagreement.
And clearer disagreement is not a problem for governance.
It is one of the raw materials governance needs.
Internal negotiation should not be designed to make the organisation sound more agreed than it is. It should be designed to make the disagreement useful before it becomes expensive.
Comments
Post a Comment