Where we started.
We were leading a 130-person company. This was how we got answers.
- Our questions pulled in two to five people.
We depended on colleagues across departments to gather the facts.
- We waited days, then stopped asking.
We asked quarterly, or never, when gathering the answer cost too much.
- We had four names for one client.
We had to reconcile about 195 companies across our systems by hand.
- We left exports everywhere.
We shared sheet links and forwarded attachments; the copies outlived the question.
- Our table said three; chat said fifteen.
We found 15 open recruiting processes and 44 candidates in one chat post while our official table sat untouched since March.
- Our chat AI could not see our records.
We still gathered everything by hand and pasted it in.
- We did not know our own contracts.
We believed every agreement signed after one date used the new template; the review disproved it.
- We founders were the decision bottleneck.
Our retrospective also named “not enough horizontal communication”.
Hardly anyone works this way yet.
- 1.5 h
- a week on average among senior executives who use AI regularly (NBER).1
- 9 in 10
- executives report no effect on their firm's employment or productivity (same NBER survey).1
- 1 in 4
- enterprises had not enabled connectors to company data in OpenAI's customer telemetry.2
- 29%
- of surveyed knowledge workers have AI embedded in their flow of work, though 85% use it (Atlassian, 2026); counted in 015 as well.3
- 16%
- of AI initiatives scaled enterprise-wide; half of CEOs call their technology disconnected and piecemeal (IBM, 2025).4
Our hypothesis: single tasks leave questions that cross departments unanswered, because a chat session without our records has to be told everything, every time.
Leaders already use AI; frequency alone does not tell us whether their records are connected.5
What one question costs us: by hand, with a chat AI, and with our second brain.
By hand: whom we asked, how long we waited, what we exported and read, where a silo stopped us. By hand with a chat AI: what most leaders have today — faster reading, the same gathering and the same walls. With our second brain: the assistant queries relevant tools through own logins; missing access stays missing and we decide. Five of our questions; switch and count.
Illustrative routes · planning estimates
Not measured before-and-after results. Each comparison covers the same question and workload; it does not establish a business outcome.
Switch the route · the map below redraws
Composed example · combined note still proposed · every week
By hand: 8.2 h estimated human work for this scenario; 36 h elapsed, including waiting. 2 role groups contribute across 5 distinct source labels. People gather and check the records.
One person walks from source to source
- 1 / 9ask15 min work
Three delivery managers: what is open with your clients?
From: three delivery managers→ their lists of open items
Wait: 24 h - 2 / 9read2 h work
40 client conversations, two chats
Slack→ who is waiting
- 3 / 9read30 min work
12 mail threads
Gmail→ unanswered mail
- 4 / 9read1 h work
Notes of 12 recurring client meetings, for promises
Fireflies→ what was promised
- 5 / 9ask10 min work
Finance: who pays slower than usual?
From: finance lead→ a promise of a list
Wait: 24 hConstraint: Accounting is finance's system; you cannot open it yourself
- 6 / 9export30 min work
Payment pace list
accounting system→ a .csv
- 7 / 9export30 min work
Hours trend per client
PirxeyOS→ a .csv
- 8 / 9merge45 min work
One sheet, four sources, client names reconciled
Google Sheets→ one table
Constraint: The same client has different names in four systems
- 9 / 9meet2.5 h work
Thirty minutes, five people
→ who calls whom
Evidence → human decision
Who gets a call today; what to promise.
The account lead decides whom to call; unavailable payment records stay unresolved.
three delivery managers · finance lead
Assumptions and all routes
Minutes are planning estimates. Human work includes review; machine time is separate. Setup is not included in these routes. Sources count distinct display labels, including repositories and public records, not integrations. Repeated Slack or Gmail instances count once. Role groups are not headcounts.
Connected copies, exports and waiting are not audited; unknown is not zero. Restrictions and checks remain. A manual route can also be well organised. The chat route assumes no connected tools and illustrative reductions in reading and merging time.
| Question | Route | Human work / scenario | Elapsed / scenario | Source labels | Role groups |
|---|---|---|---|---|---|
| Which clients are waiting on us, and what did we promise them? | By hand | 8.2 h | 36 h | 5 | 2 |
| Which clients are waiting on us, and what did we promise them? | By hand + chat AI | 5.8 h | 36 h | 5 | 2 |
| Which clients are waiting on us, and what did we promise them? | With our second brain | 30 min | 45 min | 6 | 2 |
| We have a new offer. Who should hear about it first? | By hand | 22 h | 5 days | 4 | 1 |
| We have a new offer. Who should hear about it first? | By hand + chat AI | 12 h | 5 days | 4 | 1 |
| We have a new offer. Who should hear about it first? | With our second brain | 3 h | 8 h | 7 | 1 |
| Are all our contractor agreements on the right template? | By hand | 4 days | 4 weeks | 1 | 1 |
| Are all our contractor agreements on the right template? | By hand + chat AI | 37 h | 4 weeks | 1 | 1 |
| Are all our contractor agreements on the right template? | With our second brain | 17 h | 3 days | 3 | 1 |
| Do the bank transfers match our invoices, and whom do we chase today? | By hand | 9.9 h | 4 weeks | 2 | 1 |
| Do the bank transfers match our invoices, and whom do we chase today? | By hand + chat AI | 9.2 h | 4 weeks | 2 | 1 |
| Do the bank transfers match our invoices, and whom do we chase today? | With our second brain | 1 h | 4 h | 3 | 1 |
| Did we log last month properly, and who is actually free next month? | By hand | 9.8 h | 3 days | 2 | 2 |
| Did we log last month properly, and who is actually free next month? | By hand + chat AI | 7.5 h | 3 days | 2 | 2 |
| Did we log last month properly, and who is actually free next month? | With our second brain | 1.5 h | 2 h | 5 | 2 |
For the selected connected scenario: 15 minutes of machine work, estimated separately. Five recipes we run separately; the single note is the composed version.
The questions behind our decisions.
We ran into these questions across finance, delivery and sales. Each needs different evidence. In the graph above, the same answer has three routes: by hand, by hand with a chat AI, and with our second brain. We show the hours, people and systems each depends on, then the jobs behind it.
Tick the questions you asked this month to count the records each answer needs.
Which clients need an account-owner check this week?needs the company
A chat window cannot see this
What the answer draws on
- Slack
- Gmail
- Fireflies
- accounting system
- PirxeyOS
What a chat window is missing
We need client conversations, the last meeting's promises, payment pace and hours together.
Where we would start again
Name the account owner; run the reconstruction once by hand and once with a second brain, on their own logins.
Our job and its evidence: Who is waiting on us?
Where is margin leaking, project by project?needs the company
A chat window cannot see this
What the answer draws on
- PirxeyOS
- rates
- accounting system
- ClickUp
What a chat window is missing
We need hours, rates, invoices and assignments together. We have only proposed this check.
Where we would start again
We would check one project's records by hand before comparing a second run with the second brain.
Work we have only proposed: analyses still proposed
Can we take the new project in March without hiring?needs the company
A chat window cannot see this
What the answer draws on
- ClickUp
- Slack
- company wiki
- HubSpot
What a chat window is missing
We need current assignments, sales prospects and signed hires. Our official table had fallen behind the chats.
Where we would start again
We would compare the staffing report with the table and count the conversations it needed.
Our job and its evidence: the staffing radar
Are all our contracts on the current template?a window can start it
A chat window can start this
What the answer draws on
- Drive + Docs
- the template
- operations
What a chat window is missing
We can paste two agreements for comparison. Our full review also needed missing files and unreadable scans identified.
Where we would start again
We would check our assumed template changeover against signed files; ours proved wrong.
Our job and its evidence: The filter hid it
Who should we chase for payment today?needs the company
A chat window cannot see this
What the answer draws on
- Gmail · bank notification emails
- accounting
- Google Sheets
What a chat window is missing
We need bank receipts beside invoices, including payments under another name and partial payments.
Where we would start again
We would count every place we record a payment; updating one left reminders elsewhere.
Our job and its evidence: Who actually paid?
We built something new. Who should hear about it first?a window can start it
A chat window can start this
What the answer draws on
- HubSpot
- LinkedIn + Sales Navigator
- Gmail
- Slack
- Fireflies
What a chat window is missing
We need earlier client questions and our colleagues' recent contact before deciding whom to write to.
Where we would start again
We would put the source beside each name and check whether a colleague had already contacted them.
Our job and its evidence: the launch list
Is this quarter's work actually serving the objective?needs the company
A chat window cannot see this
What the answer draws on
- strategic repository · goals
- standard repository · weekly notes
- HubSpot
- Google Calendar
What a chat window is missing
We need our objective beside departmental evidence for and against it. This comparison is still proposed.
Where we would start again
Put the objective on a page its intended readers can access; then ask each department's note for evidence against it.
Work we have only proposed: analyses still proposed
Did the sales promise survive the handover to delivery?needs the company
A chat window cannot see this
What the answer draws on
- HubSpot
- Fireflies
- ClickUp
What a chat window is missing
We need the sales-call promises beside delivery's agreed scope. We have only proposed this comparison.
Where we would start again
We would list one project's sales promises beside its delivery tasks and check the gaps.
Work we have only proposed: analyses still proposed
How much of our work is actually AI-assisted?needs the company
A chat window cannot see this
What the answer draws on
- PirxeyOS
- GitHub
- ClickUp
- PirxeyOS · licence usage
What a chat window is missing
We have survey replies. Comparing work records and licence use remains a proposal; neither proves an effect.
Where we would start again
We would compare team totals and investigate discrepancies without treating usage as proof of productivity.
Work we have only proposed: our survey; the records comparison remains proposed
How should we explain the reorganisation to the board?a window can do it
A chat window can do this
What the answer draws on
- our own notes
What a chat window is missing
We can supply the notes ourselves. A chat session has what it needs for a first draft.
Where we would start again
We would keep using a chat session for this draft and review what it says.
What do we need from this 40-page report before the meeting?a window can do it
A chat window can do this
What the answer draws on
- the report
What a chat window is missing
We can supply the whole report; the summary needs no other company records.
Where we would start again
We would summarise the report and check the claims that affect our decision.
What does the new regulation mean for companies like ours?a window can start it
A chat window can start this
What the answer draws on
- public sources
- our own contracts and processes
What a chat window is missing
We can read the public regulation. We still need our contracts and processes to establish where it applies.
Where we would start again
We would compare the general explanation with our own files and have a person check its application.
Our job and its evidence: reading everything, not a sample
We use “needs the company” for answers requiring our own records, and “a window can do it” for tasks supplied with one document. “Inputs listed” counts distinct source labels under the selected questions.
What works, and what still needs us.
| Need | In our work | Still ours to do |
|---|---|---|
| Work through own logins | A missing source stays marked as missing. | Review access. CRM administration still has a shared key; hosted models process the text we send.6 |
| Put records together | Separate skills check replies, promises, payments and hours. | Build and verify the combined note; it remains proposed. |
| Check the evidence | A person caught a wrong hours match and a filter hiding a changed clause. | Check the source, the match and the comparison method. |
| Look beyond a sample | The agreement review challenged our assumed template changeover. | Resolve missing files and unreadable scans. |
| Repeat useful work | Payment matching runs on a schedule; doubtful matches go to finance. | Count setup and review. The first month can cost more than manual work.7 |
One question to try again.
Question: Which clients are waiting on us?
Owner: The account lead.
Records: Slack, Gmail and Fireflies, through own logins. Payment information stays missing unless an authorised person can provide it.
Compare: Reconstruct the same period manually and with the assistant. Keep setup, human review and machine time separate. Record missed promises and unresolved sources.
Decide: Did the evidence change whom we should call or what we should promise? Choose the next question after reviewing this one.
A proposed trial, not a completed experiment. Read the example in 020.
Where our numbers come from.
Executive use and reported effects. NBER, Firm Data on AI, February 2026. About 6,000 executives across four countries. Regular AI users report 1.5 hours weekly on average; nine in ten surveyed executives report no employment or productivity effect. Self-reports, not audited results; the usage figure concerns the regular-user subgroup.
Connected company data. OpenAI, The state of enterprise AI, 2025. Roughly one in four enterprises in the vendor's customer telemetry had not enabled company-data connectors. That customer population is not all enterprises; it does not prove our hypothesis.
Use versus workflow. Atlassian, State of Teams 2026. The 85% and 29% figures concern 12,035 surveyed knowledge workers; 173 executives were also surveyed. January–February 2026. Vendor research, self-reported. Also discussed in 015.
Scale and fragmentation. IBM, 2025 CEO Study. Estimates from 2,000 CEOs in 33 countries: 16% of AI initiatives scaled enterprise-wide; half of respondents described disconnected technology. February–April 2025. Different denominators: initiatives and CEOs. Vendor research, not audited outcomes.
Leaders already use AI. Microsoft, Work Trend Index 2025. Survey of 31,000 workers across 31 markets: 69% of leaders and 45% of employees report regular use. Separate respondent groups; vendor research. Frequency does not establish connected use.
Our data boundaries. 017: access and hosted processing. Own logins do not keep model processing on the laptop. Our setup description is not a security certification.
Planning estimate, not measured payback. Payment matching: 12 hours setup plus 1 hour monthly human review gives 13 hours in month one, versus an estimated 8.8 hours manually for the same workload. Machine time is separate. Maintenance remains additional work.