One room to establish the facts, resolve the exceptions and commit to what happens next. Built for IT services firms and financial technology providers.
Tacit focuses on the high-stakes work before delivery: finding what is true, defining what will change, and getting the right people to commit.
Fixed-fee or outcome-led engagements break when scope, baseline and success measures live in scattered conversations. Put a reusable, AI-assisted discovery and commitment system under every engagement.
Each new bank or credit union introduces policies, systems, people, exceptions and sign-offs. Use a purpose-built Mission Room to turn this complexity into confirmed configuration decisions and an implementation-ready package.
Every Mission Room follows the same spine. Tacit structures the work needed to reach a confirmed handoff. Your existing implementation, integration and verification teams remain accountable for delivery.
Agents interview stakeholders and inspect source material.
Draft registries, configuration inputs, evidence and gaps.
Leads resolve gates and confirm each critical decision.
Configure, integrate, migrate and implement.
Measure results and run go-live assurance.
The output: an implementation-ready package, including confirmed requirements, configuration parameters, evidence, open issues, accountable owners, and success criteria. Not a replacement for the system of record.
Provider: a Third-Party Risk Management (TRM) SaaS vendor.
Client: the credit union it is onboarding.
Use case: the discovery phase of a TPRM SaaS rollout. Establish the client's real vendor landscape (who they depend on, how critical each is, what data is exposed), then confirm the decisions the platform will be configured with, so the client is implementation-ready.
The client's material seeds its playbook. The Lead defines gates before BATTUTA begins discovery.
BATTUTA interviews relationship owners, reveals shadow vendors, and escalates contradictory answers.
E-signature service used for lending; owners disagree on fallback processes.
“If the e-signature service were unavailable for a week, how would a loan close?”
Artifact agents assemble draft evidence and configuration inputs. The Lead reviews before commitment.
The Lead confirms records or overrides BATTUTA's draft assessments. Exceptions cannot be silently approved.
Resolve the e-signature service's criticality with the supporting evidence in view.
| Vendor | Tier | Status |
|---|---|---|
| Core processor | Critical | Review |
| E-signature service | Pending | Escalated |
| Card processor | Critical | Confirmed |
| Payroll provider | High | Confirmed |
Illustrative example only. Figures and client name are examples; no production workflow, decisions, or client records are changed.
This is not a chatbot on top of project management. It is a reusable operating system for the first consequential decisions of an engagement.
Expertise, interview plans, decision gates, data models, integrations and artifacts designed for the work, not a blank workspace.
Transcripts and source documents are attached to drafted findings. Critical decisions reach a named human, with overrides recorded.
Reusable methods and templates improve across engagements while confidential client material remains separated and controlled.
We're building for firms whose reputation depends on confidentiality, defensible decisions and human accountability.
Tacit Chassis is described as Apache-2.0 open source in the current research materials; specific fork licensing and implementation availability should be verified before deployment.
As services become more software-driven, the missing infrastructure is neither a personal copilot nor a fully autonomous services provider. It is the space between them: expert-led work on an agent-assisted, reusable mission chassis.
This is the thesis. The flagship TRM mission is a concrete way to test it. We distinguish early ideas and experiments from demonstrated production outcomes.
Read the research below ↓“We can know more than we can tell.”
Michael Polanyi, The Tacit Dimension (1966), p. 4
A bounded workspace for expert engagements between a provider and its client. Agents draft. One named person stands behind every finding.
Engagements run on meetings, live and recorded, by web or phone, one-to-one or in groups, and on the slides and diagrams that follow. What matters is said in them and never recorded as a decision. Weeks later no one can say which conversation a conclusion came from, who confirmed it, or where the client's material sits. Agents draft in minutes, so the gap grows.
One client, one mission, a start and an end. Inside, the Judgment Work pattern runs: agents draft, a named Lead confirms, and the mission commits. The client's context and the provider's method stay on separate shelves. We hypothesise both sides will trust it more than meetings, chat and file shares.
From paper to open source. The Mission Room ideas in the paper are being built as open-source software under Apache-2.0, free to fork for your own kind of work. The paper sets out seven hypotheses we want to test with partners. We are looking for academic feedback, industry collaboration and design partners.
Five properties from the paper. About 60 seconds. Nothing is stored.
One real mission, one client. Not a pilot of software; a test of the idea.
Free for the first design partners. We will not name you or your client without written permission. Your data and IP stay yours (below).
What we may do with design-partner data: only what you consent to, in aggregate, for research. Never model training. Never named without written permission. The optional control plane sees metadata only, never mission content.
Tell us a little. A person replies, usually within 48 hours. No client names, please.
In short. Expert engagements run on digital meetings, live or recorded, and on the slides, diagrams and chats that follow. None of these record who confirmed what. Agents now draft findings in minutes, so "who decided" gets blurrier. We propose a workspace per engagement that runs the Judgment Work pattern: agents draft and one named person confirms. We want to test whether you and your clients would trust it.
You interview a dozen people at a client, on calls and in meetings. You read what documentation exists. You form a view and write it up. Weeks later, someone asks which conversation a conclusion came from, which findings you examined yourself, and where the client's confidential material is now. Most firms can't answer quickly.
Even as offices refill, knowledge work stays digital: web and phone meetings, one-to-one or in groups, live or recorded, with chat and shared files in between. The whiteboard has become Figma or draw.io, and now an AI draws the diagram. The work is richer than ever, and the judgment inside it is no easier to trace.
The most valuable part of the work is the part no tool records.
Agents can interview, transcribe and draft. In a field study of 758 consultants, access to a model improved speed and quality on tasks inside its ability, and made people less likely to get the right answer on a task just outside it. The edge of the model's ability isn't visible to the person using it.
That removes much of the junior work, and it raises a harder question than speed: when an agent drafts a finding and a person forwards it, who decided?
Who owns what? What the client's people said belongs to the client. Your method belongs to you. Where does it live, and who can reach it? Including which model services see it. How can you prove it? In the language their security team already uses.
If a workspace can't answer these in the system itself, not just in a contract, a careful client won't let an agent interview its people.
Judgment work is work where you are paid to draw out what other people know and then put your name to a verdict: a due-diligence finding, a readiness assessment, a migration plan. It has the same shape every time, and agents make that shape worth writing down. We call it the Judgment Work pattern. It has three moves, which the paper names Draft, Confirm and Commit.
Here is one record, with invented content, moving through the pattern:
The agent proposes; the person decides. No fork of the software can remove the Confirm step.
If agents absorb the junior tasks, the staffing pyramid thins. We describe a smaller team: a Lead who is accountable and the only one who confirms, and an optional Operator who runs the room. The open worry is apprenticeship. Juniors used to learn judgment by doing the work the agents now do.
This is a position paper. We have no field data. We state eight hypotheses, for instance how often leads override an agent's score, and whether interviews change conclusions that documents alone would reach. If the answers are "rarely" and "no", the idea is weaker than we think, and we will say so.
We are looking for five to eight practitioners to run a single engagement in a Mission Room. In return: a co-authored case study and a say in what gets built. Writing a playbook needs your expertise, not code.
Describe your missionRead the full paperResearch supported by Theoremlabs.io, which also develops the open-source Tacit Chassis (in development, not yet public).
Accountability, Ownership and Data Sovereignty in Agent-Assisted Professional Services
P. Sarode · R. Nair · Adjunct Researchers-in-Residence, Future of Ways of Working, Theoremlabs.io
Expertise engagements between a provider and its client depend on knowledge nobody has written down: expert judgment, and what the client's people know but have never documented. They run on chat, file shares and data rooms, which record neither who knows what, who owns what, where it resides, nor who decided. AI agents can now interview, draft and score at scale, which makes those gaps more costly. We propose the Mission Room: a bounded, staged workspace in which agents draft and a named person confirms, the client's context remains the client's, and the provider's expertise remains the provider's. We argue that such a workspace will not be taken seriously unless ownership is separated and enforced by the system, and unless its answers to data sovereignty, security and privacy can be demonstrated under assurance standards such as SOC 2. We specify the Mission Room through these preconditions, a governance pattern called Draft, Confirm, Commit, a consolidated set of invariants, and an operating model in which a Lead and an optional Operator work with an agent roster in place of the traditional pyramid. We describe an open-source reference implementation, Tacit Chassis, built to hold the invariants, and state eight testable hypotheses. The paper proposes a design and reports no field data.
Read onlineDownload PDFSend comments
@misc{sarode_nair_2026_tacitwork,
author = {Sarode, P. and Nair, R.},
title = {Tacit Work Needs a Mission Room: Accountability, Ownership and
Data Sovereignty in Agent-Assisted Professional Services},
year = {2026},
month = oct,
note = {Working paper, version 0.8},
url = {https://tacit.theoremlabs.io/paper/v0.8/}
}
Comments, critique and collaboration are welcome at research@theoremlabs.io. Each posted version keeps its own address; this page links to the latest.
Corporate knowledge management stored what people had written down. The part that matters most, what experts know and decide, was never managed. Agents can now interview, draft and score, and a named person can confirm. That makes a system for judgment management possible, and knowledge management ripe for rebirth.
From paper to open source. The Mission Room ideas in our working paper are being built as open-source software under Apache-2.0, free to fork for any kind of work. We are seeking design partners for the first missions, academic feedback and industry collaboration, and a few investors who see the category early.
Every generation of enterprise software gave one kind of asset a system of record. Judgment, meaning who knows what, what an expert decided on what evidence, and who stood behind it, is still managed in meetings, chat, shared files and slides.
| Era | What it managed | The system of record |
|---|---|---|
| 1990s, knowledge management | Documents and written-down know-how | Intranets, document and knowledge repositories |
| 2000s to 2010s, enterprise software | Customers, money, people, work | CRM, ERP, HR systems, issue trackers |
| 2020s, agents | Explicit traces and memory inside one enterprise | Knowledge bases, context graphs, decision logs |
| Next | Judgment across a provider and its client: who knows what, what was decided, on what evidence, confirmed by whom | A system for judgment management (the Mission Room, proposed) |
Lineage: Polanyi (1966) on tacit knowing; Nonaka and Takeuchi (1995) on converting tacit knowledge; Davenport and Prusak (1998) on managing what organizations know. Full references are in the working paper.
A Mission Room runs the Judgment Work pattern (Draft, Confirm, Commit) in a governed, time-boxed engagement between two organisations: agents draft, a named Lead confirms, client context stays the client's, provider method stays the provider's. The output is a scored, confirmed decision record with provenance to the source interview. That is a system of record, not a knowledge base. Decision logs and context graphs record what an enterprise does inside its own walls; a Mission Room records expert judgment across the boundary between two organisations.
The same base serves any provider that must understand how a client really operates before delivering. Two shapes the paper sketches for financial institutions (illustrative, not built):
| Mission | Provider → client | What the Lead confirms | |
|---|---|---|---|
| Customer onboarding discovery | Fintech implementation team → bank or credit union | Each process step or control gap, owner, severity, before go-live | illustrative |
| Third-party risk go-live | RegTech vendor → regulated firm | Each vendor relationship, evidence, readiness | illustrative |
| AI adoption assessment | Advisory firm → enterprise | Each AI use case, owner, controls, score | specified |
Implementation and onboarding are where B2B fintech vendors lose time. A Mission Room is a candidate for the discovery stage of that motion.
Adoption first. Consulting firms and vendors will not put their method into a closed tool owned by a competitor. Apache-2.0, a published playbook format and a rule that no fork can remove the human confirmation gate make the Chassis a standard others can build on. Honest caveat: open-source users will be builders, not consultants, so the hosted route has to serve the rest.
Seven public hypotheses. Three that matter most to a buyer:
| ID | If this holds… | …the case strengthens; if not, it weakens |
|---|---|---|
| H1 | Leads override a stable, measurable share of agent scores | Evidence that the human gate is real value, not ritual |
| H2 | Missions reach Commit faster than comparable work on general tools, at no loss of quality | The efficiency case |
| H3 | Interviews change conclusions that documents alone would reach | Why elicitation is the product |
A 30-minute conversation to pressure-test the thesis. For investors with bank and vendor networks, we would value introductions to implementation teams and institutions that could run one real mission as design partners.
Book 30 minutesRead the paperResearch pages keep the paper's own typography so a reader arriving from a citation sees the same document. Everything around them is the Tacit shell.
For an IT services firm, bring an outcome-based engagement. For a fintech or RegTech vendor, bring a complex new-customer onboarding. We'll explore the process, evidence and human decisions that could become a reusable Mission Room.
Initial contact by email; no customer data or confidential client names needed.