The approach
Find the use case that eats the most hours. Build there first.
Most AI work in a mid sized company starts with a tool and goes looking for a use case. We start with a decision somebody is accountable for and time the work sitting in front of it.
The idea
Every decision has a line in it
Above the line is assembly: reading the invitation, checking history, pulling margin, confirming capacity, writing the summary. It is repeatable, timeable, and mostly retrieval. Below the line is the call itself, which takes minutes and depends on judgment somebody is paid for.
We automate everything above the line and nothing below it. The person who owns the outcome still makes the call, only now it arrives prepared, with the evidence attached, instead of after forty minutes of digging.
The note written at the line, what the decider needs in front of them, is the build specification. Everything the system produces exists to satisfy that note and nothing more.
- Above the line
- Retrieval, extraction, classification and summarizing across messy sources. This is what current AI is good at.
- Below the line
- Weighing tradeoffs someone is accountable for. This is where models are least trustworthy.
We automate up to the line and stop. The accountable person gets the evidence already assembled, with a citation on every fact, and the call stays theirs.
- Read the invitation and addenda
- Check history with the GC
- Check backlog and crews
- Pull margin on the last three like it
- Check bonding and prequal
Below the line · the call
Bid or no bid
39 min
assembled by AI
90 sec
deciding, still a person
Measuring before building
Two ways we measure, and they answer different questions
Every engagement measures before it builds. Which measurement you need depends on what you already know, which is why neither one comes first.
The AI Adoption Baseline shows where to look
Conversations are the method. We sit with leaders and their teams with no manager in the room, inventory what is actually running and on whose account, and score how aligned the picture is. The rollup names the decision lines we heard about along the way, so it ends by pointing somewhere. It is not an audit and never a hunt: we report patterns, not people.
Line Mapping measures those lines
The stopwatch is the method. We interview the people who own and prepare each decision separately, watch real instances being prepared and time them step by step, then send the numbers back to the people who gave them. That validation pass is what makes every figure defensible for a year.
An illustrative set of line maps, plotted
Scoring
Every line, placed by cost and feasibility
Each line is scored on six weighted dimensions. Those scores plus the annual assembly cost place every line in one of four quadrants. The do not build list gets drafted first, so every recommendation has to survive it.
- Bid or no bid: Build first
- Equipment, rent or buy: Do not build
- Which growth spend is returning: Fix first
- Warranty claim triage: Batch or skip
- Which leads are worth an hour: Build first
- Trade show spend: Do not build
- Which GCs to pursue: Fix first
- Subcontractor selection: Batch or skip
- Change order pricing: Build first
Build first
Two or three at most. These go into the sprint proposal with what a working version looks like.
Fix first
The value is real and something is in the way, usually data or ownership. We name the fix, not a build.
Batch or skip
Group several into one sprint or leave them. Never sold on their own.
Do not build
Named in the report with the reason in full sentences. This list is why the rest is believed.
- Data reachability
- Output acceptance
- Trigger clarity
- Variability
- Owner pull
- Error tolerance
See a sample report from a finished Line Mapping engagement.
Principles
Four things you can hold us to
01
You get a number before you get a proposal
We scope from timed steps your own people supplied, and we publish the arithmetic so you can argue with it.
02
The judgment stays with the person accountable for it
The method has a named stopping point. We automate the gathering, checking and summarizing, and the call stays with whoever owns the outcome.
03
It runs on your own software, so nothing turns off when we leave
Live systems run on your company's own Microsoft, Google or AI subscriptions.
04
You will not spend the first meeting explaining your industry
We have run B2B marketing across manufacturing, distribution, construction and logistics. We know how a contractor pursues work.
Questions
Before you ask
What do you mean by the decision line?
Every recurring decision has assembly above it and a judgment call below it. Above the line is retrieval, extraction, classification and summarizing across messy sources, which is what current AI is good at. Below the line is weighing tradeoffs someone is accountable for, which is where models are least trustworthy. We automate up to the line and stop.
Why not automate the decision itself?
Because somebody is accountable for the outcome and a system cannot hold that. We make the call cheap to make, well evidenced and fast, with a citation on every fact, and the person who owns the result still makes it. If a decision genuinely has no judgment in it, it was a lookup rather than a decision and it did not need us.
How do you decide what is worth building?
Each line is scored on six weighted dimensions: data reachability, output acceptance, trigger clarity, variability, owner pull and error tolerance. Those scores plus the annual assembly cost place the line in one of four quadrants. Data reachability and output acceptance can sink a case on their own.
What does the do not build list actually do?
It gets drafted before the recommendations, so every build we propose has to survive it. Mapping twelve things and recommending nine is the most common failure in this kind of work, because recommending everything is recommending nothing. The list is why the rest of the report is believed.
Do you need access to our systems before you can score anything?
No. The scoring comes out of conversations and timed walkthroughs, not out of an integration. We do ask where the data lives and whether anybody can already export it, because data reachability is a hard constraint and a line blocked on it goes on the fix first list rather than the build list.
Why watch the work instead of just asking?
Because people are reliably wrong about how long their own preparation takes, almost always low. Ask an estimator how long a bid or no bid decision takes and they will say a minute or two, which is true of the decision and not of the forty minutes in front of it. Timed walkthroughs are where the honest numbers come from.
Talk to us
Find out what your decisions really cost
Thirty minutes on a call, or about ten minutes on your own with the assessment. Either way you leave with a number.