Skip to main content

Product note · Mission Delivery Notes

When building gets cheaper, agreement becomes the work.

AI is cutting the cost of producing software. It is not cutting the cost of getting a contracting officer, a program owner, a security team, an identity team, and an operations team to agree. That is where government projects now stall. We have begun implementing Parley to change how that agreement gets made, starting with project management and acquisitions management.

By EVP

Published

Read10 minutes

Building is getting cheaper

For most of my career, the expensive part of a government software project was making the thing. You needed people to write the code, test it, document it, and keep it running. Every other cost hid behind that one.

That is changing. With AI in the loop, a small senior team can draft code, tests, documentation, and deployment scripts far faster than before. The evidence is not uniform. A 2025 randomized trial by METR found that experienced open-source developers took 19% longer on tasks in their own repositories when they used early-2025 AI tools. Still, the direction is clear: Google's 2025 DORA research found that about 90% of respondents use AI at work, and the cost of producing a working draft keeps falling.

When one cost falls, the next one becomes visible. On government projects, the next one is agreement.

Communication was never a small fraction

Project managers have measured this for years. In 2013, PMI's Pulse of the Profession report on communications found that for every $1 billion spent on projects, $135 million was at risk. Of that, 56%, or $75 million, was at risk because of ineffective communication. Organizations with highly effective communicators delivered 71% of projects on time, compared with 37% for organizations without them.

Government evidence points the same way. When GAO studied seven federal IT acquisitions that succeeded, the only factor cited by all seven was that program officials were actively engaged with stakeholders. Involving users in requirements, prioritizing requirements, and communicating regularly with the contractor followed close behind. Sufficient funding was cited by three.

None of this is new. Fred Brooks showed in 1975 that when every part of a task must be coordinated with every other part, the effort grows as n(n−1)/2. Melvin Conway observed in 1968 that organizations produce designs that copy their own communication structures. The communication cost was always there. It was just smaller than the cost of building.

Coordination paths multiply

When everyone needs to coordinate directly with everyone else, adding people creates more possible paths than the headcount suggests.

2 participants / 1 possible path

4 participants / 6 possible paths

8 participants / 28 possible paths

Possible paths = n × (n − 1) ÷ 2

An illustrative count of pairwise connections, not a measurement of effort, time, or cost. Actual communication needs depend on how the work is organized.

Every stakeholder has a different definition of done

Here is a typical government project in its first month. Everyone involved is reasonable. Everyone wants something different, and almost everyone has a gate the project must pass.

Contract and program

  • The contracting officer wants decisions documented and traceable to the contract.

  • The program owner wants a firm delivery date.

  • The project manager wants a Jira ticket for every piece of work, so progress can be tracked and reported.

Security, privacy, and accessibility

  • The security team wants encryption at rest, and evidence that it is on.

  • The identity team wants the new agency login system, not the old one.

  • The privacy office wants to know what personal data moves, and where it goes.

  • The Section 508 team wants accessibility conformance tested and documented before release.

Architecture and engineering

  • Technical leadership wants a design review before major work starts.

  • The technical review board wants the design to fit agency standards and approved technologies.

  • The architects want their patterns followed and a say in the tradeoffs.

  • The engineering leads want a code-quality gate, such as a minimum test coverage threshold in SonarQube.

  • Someone wants documentation in a specific format, at a specific depth, for a specific audience.

Operations and release

  • The operations team wants to know how disaster recovery works and where the backups live.

  • Change management wants every production change scheduled, reviewed, and approved by its board.

  • The communications office wants advance notice and plain-language release notes for end users.

  • The end users want a tool that fits how they actually work.

Every one of these asks is legitimate. Together they overlap, compete, and sometimes contradict each other. A firm date and a dependency on a login system that is still rolling out may not both be true. A change board that meets every two weeks may not fit a release plan built on weekly releases. A coverage threshold, a documentation standard, and a ticket for every task all draw on the same small team's time. Some requirements are missing entirely, because they sit in contract language or agency policy that nobody in the room has read lately.

Holding all of this in one head is close to impossible. Today, a brave project manager tries anyway. They chase each stakeholder, translate between groups, keep the list current, and catch conflicts when they can. Good project managers are a big reason government projects work at all. But they are people. They have other meetings, take leave, change roles, and cannot watch every thread at once. So conflicts still surface late, often after work has been built on the wrong assumption.

The coordination job is not too small for a person. It is too constant for one.

Cheaper building makes the problem sharper

When building takes most of the schedule, slow agreement hides inside it. When building gets fast, slow agreement becomes the schedule.

Research on AI-assisted delivery supports that reading. Google's 2025 DORA report concluded that AI amplifies what is already there: teams with loosely coupled systems and fast feedback see gains, while teams constrained by slow processes see little or no benefit. A year earlier, DORA estimated that rising AI adoption came with a 1.5% decrease in delivery throughput and a 7.2% reduction in delivery stability. Atlassian's 2025 developer experience survey found that developers spend about 16% of their time coding, and that half of them lose 10 or more hours a week to inefficiencies such as finding information and working across teams.

Federal process shows the same pattern from another angle. GSA reports that its FedRAMP 20x approach cut authorization time for participating cloud services to about five weeks, down from more than a year. The technology did not get ten times better. The process around it changed.

GAO still lists IT acquisition and management as a high-risk area, and the federal government invests more than $100 billion a year in IT. As AI makes building cheaper, more of that investment will depend on how quickly people can agree on what to build.

Give every stakeholder an advocate

AI already sits at the input and output of most people's work. People draft with it, summarize with it, and ask it questions before they ask a colleague. That points to a different approach to project communication.

Instead of routing every need through a human intermediary, each stakeholder sets up an AI advocate. The stakeholder tells their advocate, privately and candidly, what they need, what they can trade, and what they cannot accept. The advocates then work the problem together. They compare goals, look for overlap, and flag where two goals cannot both be true.

Each advocate holds context that the other agents have not yet seen: a constraint, a priority, or a possible trade. When an agent shares new information with the group, something previously held privately becomes part of the shared conversation. Other advocates can react immediately using the context they already hold. Negotiation can move at machine speed instead of waiting for the next meeting or reply. Stakeholder control over what is shared remains essential.

This does not replace the project manager. It gives them help that keeps working when they cannot, and a current picture of every position, overlap, and conflict.

Private, candid input is not a new idea. The Delphi method, developed at RAND and published in 1963, uses anonymous responses and repeated rounds of controlled feedback so experts can state their real positions without the pressure of a room. People are often more honest about constraints when they are not saying them in front of the group those constraints will disappoint.

From private goals to a proposed plan

Parley is in development. This diagram shows the workflow we are building toward, not a claim that every capability is complete.

  1. Stakeholders state their goals

    Each person privately records needs, constraints, and tradeoffs.

  2. AI advocates compare positions

    New information enters the group. Advocates react using context they previously held privately, exposing overlaps, conflicts, and missing requirements.

  3. A proposed plan takes shape

    Agreed work and unresolved decisions are separated for review.

  4. People decide what proceeds

    Stakeholders resolve conflicts. Government officials retain required authority.

Unresolved conflicts return to the stakeholders. The plan works alongside existing AI tools and documents.

What Parley is designed to do

Implementation of Parley has begun. We are starting in project management and acquisitions management, while the underlying pattern can apply wherever participants need to reconcile different goals and privately held context. This is the workflow we are building toward:

  1. Private intake. Each stakeholder records goals, constraints, and priorities in their own space, or uploads the documents that already define them. They decide what is shared word for word and what their advocate represents on their behalf.

  2. An advocate for each perspective. An AI agent represents each stakeholder's goals in the project conversation and keeps arguing for them.

  3. Contract and policy context. Parley reads the contract language and applicable requirements, so obligations no stakeholder raised still show up.

  4. Explicit conflicts. When two goals cannot both be true, Parley names the conflict, the stakeholders involved, and the options, instead of quietly averaging them away.

  5. Missing pieces. Parley flags gaps, such as a backup location no one has decided or a contract requirement nobody has claimed.

  6. A proposed plan. Once every stakeholder has contributed, Parley drafts a plan that shows what each decision trades off and why.

  7. People decide. The plan is a proposal. Government officials keep their decision authority, and the delivery team stays accountable for the result.

Most of a project's requirements do not conflict. They need to be found, written down, and honored. Parley is designed to move through that work quickly: confirm what is required, record what everyone already agrees on, and keep it current. People spend their attention on the conflicts and trade-offs that need a decision, not on a status meeting about everything else.

Parley does not need to be the only AI in the room. Stakeholders can work with whatever AI tools they already use, draft their position there, and bring it in. They can also upload the documents that define their baseline, such as a security plan, an architecture standard, a change calendar, or the contract itself. Parley's job is to reconcile those inputs, not to replace them.

Easy things should be easy, and hard things should be possible. That is the bar for Parley. Agreed and required work should flow on its own. Real conflicts should become possible to resolve, with the right people looking at the right question.

What has to be true for this to work

The research is candid about the hard parts, so we will be too.

  • Agents agree too easily. Recent studies of multi-agent AI systems found that agents tend to reinforce each other and settle on premature consensus. An advocate that gives in to keep the peace is useless. Parley has to preserve disagreement until a person resolves it.

  • Conflict detection is incomplete. One study that paired a language model with a formal solver flagged conflicts with high precision but found fewer than a third of them. Another found that capable models often detect conflicting instructions but rarely tell the user. Parley has to show its work and treat silence as unknown, never as agreement.

  • Privacy has to be real. Candid input only works if people trust where it goes. Stakeholders need to control what is shared and see what their advocate said on their behalf.

  • Contract text is not the whole truth. Agency policy, informal commitments, and operating realities matter too. Parley should cite its sources and leave interpretation to the people accountable for it.

  • Decision rights stay with people. Parley can propose. It cannot commit the Government, accept risk, or change the contract.

What comes next

We are continuing implementation and testing. Our initial findings have been encouraging: in early internal tests, communication exchanges that took hours have happened in seconds. We still need to establish how well those gains hold across different situations, whether conflicts are surfaced reliably, and whether people can inspect and trust the proposed result. We will report what holds up and what does not. If the advocates cannot surface conflicts earlier and more honestly than the meetings they replace, we will say so.

If your projects stall in the space between people rather than in the code, we would like to compare notes.

Sources

  1. Project Management Institute, Pulse of the Profession: The High Cost of Low Performance: The Essential Role of Communications (May 2013).

  2. U.S. Government Accountability Office, Information Technology: Critical Factors Underlying Successful Major Acquisitions, GAO-12-7 (October 2011).

  3. U.S. Government Accountability Office, High-Risk Series: Critical Actions Needed to Urgently Address IT Acquisition and Management Challenges, GAO-25-107852 (January 2025).

  4. Frederick P. Brooks Jr., The Mythical Man-Month (Addison-Wesley, 1975), chapter 2.

  5. Melvin E. Conway, "How Do Committees Invent?", Datamation (April 1968).

  6. Google Cloud, Announcing the 2025 DORA report: State of AI-assisted Software Development (2025).

  7. Google Cloud, Announcing the 2024 DORA report (2024).

  8. Atlassian, State of Developer Experience Report 2025 (2025).

  9. Joel Becker and others, METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" (July 2025).

  10. U.S. General Services Administration, "GSA celebrates major FedRAMP milestones" (August 11, 2025).

  11. Norman Dalkey and Olaf Helmer, "An Experimental Application of the Delphi Method to the Use of Experts", Management Science 9, no. 3 (1963).

  12. Study of sycophancy and premature consensus in multi-agent debate, arXiv:2509.23055 (2025), and CONSENSAGENT, Findings of ACL (2025).

  13. Requirements conflict detection combining a language model with an SMT solver, NSF Public Access Repository.

  14. ConInstruct: conflict detection in instructions, arXiv:2511.14342 (November 2025).