Watch · 9:30–11:08 · one workflow at a time
Unmesh finishes the end-to-end sequence: connect accounting through MCP servers and APIs, move one low-risk workflow at a time, and keep the current system running meanwhile. At 9:57 he sets that against the 6 to 9 months a traditional implementation takes, the contrast this lesson ends on. Watch on YouTube ↗
Jump to: 4:03 — the Monday questions · 9:57 — 6 to 9 months to implement · 13:19 — the compliance questions a board needs
Monday, 9:04: three answers due today
R-2 is decided and the first quarterly updates are in. Your inbox has three messages. The board chair asks, "What did R-2 fund, and is it working?" Finance wants spend against budget for each grantee. The auditor wants proof of how the A-31 award to Eastgate Youth Works was decided.
The questions are ordinary. The hard part is where the answers live: in Horizon's configured system, award amounts sit in one module, reviewer comments in another, the budget workbook in an attachment and the reason for the A-31 condition in a committee member's email.
So the answer becomes a project: export, Excel, weekend. By Tuesday the numbers exist, but nobody can say which budget version they reflect.
Roles are set once and rarely tested. Answers leave by export and are rebuilt by hand. The decision rationale lives wherever the reviewer wrote it. Changing anything means asking the admin, or the vendor.
Each role sees its part of one record per applicant. Questions are asked of the record and answered with sources. Suggested scores, reviewer scores and the committee decision sit side by side, each with its author.
Give each role a defined job and a defined view
Write down who does what before you test anything. One person may hold several roles, but each responsibility needs a name. The last column is the test that should fail: a denied action tells you more about access than an allowed one.
| Role | Responsible for | May see | Test that should fail |
|---|---|---|---|
| Applicant (for example, Eastgate) | Submitting, answering clarification requests, uploading reports | Its own application, requests, award terms and reporting plan | Opening another application or a reviewer's note |
| Reviewer (4 staff, 2 community volunteers) | Scoring assigned applications, declaring conflicts, recording disagreement | Assigned applications, the rubric and anchors, suggested scores with their sources | R3 opening A-17's files or export after reassignment |
| Committee | Deciding awards, setting conditions, recording overrides | All eligible applications, reviewer scores, calibration notes | Recording a decision without a rationale |
| Program owner | Onboarding, the reporting plan and follow-up | Approved terms, conditions, reporting plan and reports | Reading committee deliberation |
| Finance | Payments against conditions, spend against budget | Award amounts, payment conditions, approved budget versions | Releasing Eastgate's second payment before the condition is met |
| AI Assistant or MCP client (Claude, ChatGPT) | Reading, summarizing and answering with sources | Only what the person asking may see | Returning a restricted note to someone not permitted to see it; changing a record |
The last row is the one teams forget. A question to the AI Assistant, or to Claude through MCP, carries the asker's permissions. If a volunteer reviewer cannot open committee notes, a question cannot open them either.
R3 sits on A-17's board, declared it, and A-17 was reassigned before scoring. Test that the reassignment removed R3's access everywhere: screen, source files, export and assistant.
Give clarifications one owner. A-44's missing fiscal-sponsor letter should produce one request, on the record, visible to its assigned reviewers, not a private email from one reviewer.
Rehearse the whole round with three records
Before the pilot counts as run, walk three records through every step they touch. Each tests a different part of the round.
Horizon example · fictional
| Record | What you test | A pass looks like |
|---|---|---|
| A-44 | Missing fiscal-sponsor letter | Held, not failed. One clarification request, answered within the window, attached to the record, then scored. |
| A-17 | Reviewer conflict | R3's conflict declared and A-17 reassigned before scoring. R3's access removed from files, exports and the assistant. |
| A-31 | Corrected assessment | Both feasibility scores kept, the calibration note attached, the agreed 3 traceable to the anchor, total 70. |
| A-31 | Committee decision | Approval, condition and decider recorded. The AI's suggested scores shown separately from the people's scores. |
| A-31 | First reporting request | The quarterly update asks for what the onboarding call agreed: unique youth enrolled (target 200 over the year), attendance at 3 of the first 4 sessions, placements at 90 days, budget against actual. |
Run each test as the role that would do it. An admin account sees everything, so it proves nothing about access. Log in as a reviewer, the program owner and Eastgate.
When a test fails, fix it and rerun that step. A failed rehearsal costs an afternoon; the same failure found by an applicant costs more.
Answer the Monday questions from the record
Now return to the three messages. Each answer should come from the record with its sources, say how much of the round it covers and name what is not yet known. The ask-the-round deep dive shows how to phrase the question, set its boundaries and check the answer. Work through it if your first answer comes back as a confident paragraph with no sources, then come back.
Board chair · "Is it working?" · Eastgate, fictional
Funded: A-31, Eastgate Youth Works, a 12-week job-readiness and placement program. Approved with a condition on the employer placement role. Source: A-31 decision record.
Q1 so far: 180 registrations. 108 of 150 eligible participants attended at least 3 of the first 4 sessions (72%). Source: Eastgate's Q1 program report, as uploaded.
Not yet known: 30 registrations have no status (not yet eligible, not started, withdrawn or unknown), so enrollment against the target of 200 is not yet a clean count. Job placements at 90 days are not due yet.
At the end of Q1 that is a useful answer: attendance is on the record, the enrollment gap has an owner, and placements are honestly unknown.
Finance's question works the same way, grantee by grantee. Eastgate's answer includes the $2,000 moved from materials to transport stipends, approved in Q1 with a reason, and reads spend against that approved budget version.
The auditor's question needs more than an answer. It needs a trail a stranger can follow: the application as submitted, both feasibility scores, the calibration note, the decision, the condition and whether it has been met. The audit trail deep dive reconstructs A-31 and Eastgate's budget change step by step. Use it to build the packet before Friday.
Run the pilot as a real cycle, with a test log
A pilot is a real round with real applicants and real money, run so every step leaves evidence. R-2 qualifies: grants of up to $25,000, 64 applications, 6 reviewers and about 10 awards. Small enough to watch closely, real enough to mean something.
Keep a test log from the first day. For each step, write what should happen, what did happen, the evidence and who owns the fix. Record the staff time each step took, too; that is the number your keep-or-extend decision will lean on.
| Step tested | Expected | Evidence to keep | Owner |
|---|---|---|---|
| Incomplete application (A-44) | Held with one clarification request | Request, answer, timestamp | Grants manager |
| Conflict (A-17) | Reassigned before scoring; access removed | Declaration, reassignment, denied-access screenshot | Reviewer lead |
| Suggested scores | Each criterion shows the passage it rests on | Three sampled applications checked against source | Reviewer lead |
| Override (A-31) | Both scores and the reason kept | Calibration note | Committee chair |
| Condition to finance | Second payment held until the placement role is confirmed | Payment record showing the hold | Finance lead |
| Report on arrival (Eastgate Q1) | 30 unresolved registrations flagged, not counted as enrolled | Check result and request sent to Eastgate | Program owner |
| Assistant access | A restricted committee note is not returned to the program owner | Question asked and the refusal | Grants manager |
| Correction | Earlier version kept when a value changes | Version history for the changed field | Grants manager |
An honest limit belongs in the log as well. One pilot shows that the workflow ran and where it broke. It does not show that the scoring is fair across every program you run, and Eastgate's attendance figures do not show that the program caused later placements. People make the decisions, every AI reading is checked against its source, and questions of cause need an evaluation design.
Decide: keep, extend or replace
At the end of R-2 you have what a demo cannot give you: proof on your own data. Use it to decide on each workflow you tested.
Keep a step that passed and that your team can repeat without the vendor in the room. Extend by adding the next workflow in the next cycle. Replace the matching piece of the configured system only after the new step has held up over more than one cycle. Anything that failed gets fixed and retested before it enters that conversation.
Traditional grant-system implementations take 6 to 9 months. For a team whose current system works, that is the wrong first bet: months of migration and retraining before one application is read differently, and a long contract signed on the strength of a demo.
One workflow per cycle is the better bet. Horizon keeps its configured system for its large grants and adds one step at a time: application review first, then onboarding, then grantee reporting, then compliance and audit, then a connection to accounting through MCP or APIs. Each step builds on the same record, so the fifth step inherits everything the first four learned.
For Horizon, the likely decision after R-2 is keep application review, extend to onboarding for the next round, and replace nothing yet. If the next move is a new program rather than a new workflow, the launch deep dive shows how to put it on the tested workflow without copying eight years of configuration along with it.
Turn your five artifacts into a vendor test script
The course has produced five artifacts: the intake specification (Lesson 1), the anchored rubric and calibration record (Lesson 2), the reporting plan for Eastgate (Lesson 3), the report-completeness rubric (Lesson 4) and this test log. Together they are a requirements brief, and a test script any vendor should pass on your data, Sopact included.
| Test | What to bring | What a passing answer shows |
|---|---|---|
| Intake that listens | Your form and three permitted past applications, one incomplete | The gap raised as a clarification on the same record, not scored as zero |
| Rubric scoring with sources | Your anchored rubric and five applications your team has already scored | A suggested score per criterion with the passage behind it; differences from your team explained |
| Override and calibration | One application your reviewers disagreed on | Both scores kept, the anchor decision recorded, the change traceable |
| Award to reporting plan | Onboarding call notes or transcript and the award terms | Measures, targets and due dates on the same record as the application |
| Report checked on arrival | A real grantee report and budget workbook | Missing or inconsistent items flagged with a request back to the grantee |
| Ask the round | Three questions your board, finance and auditor asked last quarter | Sources, coverage and what is unknown, repeatable next week |
| Access and audit | Your role table and one past decision | A denied request that stays denied through export and MCP; the decision reconstructed with versions |
Run the same script with every vendor, on the same data, with the same people watching. The comparison is then about your team's work, not the more polished demo.
In Sopact Sense, each applicant keeps one persistent unique ID from application to award to reporting. The Intelligence Cell reads each answer and document as it arrives, the Intelligence Row summarizes each applicant or grantee, and the AI Assistant answers questions with sources. Claude or ChatGPT can query the same governed data through MCP. Ask us to run this script on your own records, and hold us to the passing column.
Put it into practice
Run R-2 as a pilot and write your test script
- Open step 5 of the workbook and
horizon-r2-applications.csv. Fill in the role table for your team: who holds each role, what each may see and one test that should fail. - Rehearse A-44, A-17 and A-31 through clarification, reassignment, corrected assessment (use
calibration-scores.csv), committee decision and first reporting request. Log the expected result, actual result, evidence and owner for each. - Answer the board chair's question from
eastgate-q1-update.csvwith sources, coverage and what is still unknown. - Run one denied-access test through the assistant and one through an export.
- Mark each workflow keep, extend or replace, and name the one workflow you will add next cycle.
- Use the vendor test script template in the workbook to turn your five artifacts into a one-page requirements brief.
Download the workbook (PDF)
Practice data: horizon-r2-applications.csv · horizon-rubric.csv · calibration-scores.csv · eastgate-q1-update.csv
Check your reasoning before moving on
A-44 is held, not rejected, and scored only after its letter arrives. A-17 is reassigned before scoring, and R3 cannot reach its files through an export or a question to the assistant. A-31's 70 is traceable to four criterion scores, with both feasibility scores and the calibration note kept. Finance can see the condition, and the second payment stays held until the placement role is confirmed.
The board answer gives 72% attendance (108 of 150), flags the 30 registrations without a status and says placements are not yet known. After one cycle, Horizon keeps application review, extends to onboarding and replaces nothing yet. The test script is ready when every row names what to bring and what a pass looks like.
Ask any vendor, including us
Log in as a volunteer reviewer and try to open a reassigned application, then ask their AI assistant for it; both should refuse. Ask them to reconstruct one past decision from your records, with both reviewers' scores and the approval. Ask last quarter's board question on your data and look for sources, coverage and unknowns. Then ask what adding onboarding next cycle would take without touching the tested review step.
Questions grant teams ask
Do we have to replace our grant management system?
No. Keep it running for the programs it handles well. Pilot one complete cycle on your smallest program, alongside it, and decide from the results. Horizon keeps its configured system for large grants while R-2 runs AI-native. Replacement comes later, one workflow at a time, and only after a step has held up for more than one cycle on your own data.
Can automated ranking replace the committee?
No. Suggested scores and ranked lists prepare the meeting; the committee makes the decision and records its reasons. In R-2, two reviewers disagreed on A-31's feasibility, and calibration, not the AI, settled the anchor. What AI changes is the reading: every application read the same way, with the passage behind each score, so people spend their time on judgment.
What should we pilot first?
Your smallest, lowest-risk program: a small grant round, a fellowship or a scholarship. Look for low dollars at risk, a short cycle, clean data and a rubric the team already argues about. Run application review end to end first. Onboarding, grantee reporting, compliance and audit, and the accounting connection follow in later cycles.
How do we compare vendors fairly?
Give every vendor the same test script, the same permitted data and the same people watching. Bring your rubric, applications your team has already scored, one disputed decision, a real grantee report and three questions your board asked. Judge the answers by sources, coverage and what they admit is unknown, and by how much your staff must rebuild next cycle.
Deep dives for this lesson
Open one when the exercise raises that question, then come back. Each uses the same Horizon example.
- Ask your whole grant round anything (Assistant + MCP)Ask R-2 in plain language and get answers with sources, coverage and what is still unknown.
- How do you build a grant audit trail?Reconstruct A-31's award decision and Eastgate's $2,000 budget change from sources and approvals.
- How do you onboard a grant or RFP program?Launch the next program on the tested workflow: records, intake, review and five failure tests.