First-party data · Finance AI
79% of the finance AI workflows we built had no off-the-shelf answer
Last updated: July 2026 · counted 2026-07-20
We publish every workflow we build as a runnable freebie. That library is now 107 builds deep, which makes it a dataset: it records, build by build, whether an existing tool could do the job or whether we had to write it. Here is what it says, with the method and the limitations stated up front so you can decide how much weight it carries.
107
builds published in the public library
counted 2026-07-20
79%
had to be built from scratch (85 of 107)
no usable existing project to start from
94%
of finance builds were custom (73 of 78)
the off-the-shelf gap is worst in finance
25%
ship with a live interactive tool (27 of 107)
runnable in a browser, not just a prompt
How the builds break down by origin
Every build is tagged with where it started. Either an existing project did the core work and we wired it into a business workflow, or nothing usable existed and we wrote it. 85 of 107 landed in the second bucket.
| Origin | Builds | Share | What it means |
|---|---|---|---|
| Built from scratch | 85 | 79% | No existing project did the job. The build is ours end to end. |
| Packaged from open source | 17 | 16% | An existing repo does the core work. The build wires it into a business workflow. |
| Buildout candidate | 4 | 4% | An existing repo is a viable starting point but needs substantial work. |
| Fork candidate | 1 | 1% | An existing repo needs forking and material changes. |
The split gets sharper inside finance. Of the 78 finance and data builds, 73 were written from scratch. That is 94%, against 79% across the library as a whole.
Why the gap is worst in finance
Three patterns repeat across the 78 finance builds, and they explain why generic tooling keeps falling short.
The inputs arrive unstructured. Statements, rate confirmations, capital account letters, and data-room PDFs are the raw material, not clean API responses. A tool that assumes structured input solves the easy half of the problem.
The output has to cite its source. In diligence, close, and reconciliation work a number nobody can trace back to a page is unusable, however good the model is. Source-grounding is a build requirement rather than a feature, and it is the thing most off-the-shelf tools leave to the user.
The workflow is shaped by one firm's conventions. Two funds close the month differently, and the difference is not cosmetic. That is where the last 30 to 40% of the work sits, and it is the part a generic product cannot ship.
What domains the builds cover
| Domain | Builds | Share |
|---|---|---|
| Finance and data | 78 | 73% |
| Business operations | 13 | 12% |
| Big 4 advisory | 4 | 4% |
| Marketing automation | 3 | 3% |
| Developer automation | 3 | 3% |
| Cybersecurity | 2 | 2% |
| Content automation, desktop AI, legal, real estate | 4 | 4% |
How much technical skill each build needs
98 of the 107 need no coding. That is deliberate: a build a finance team cannot run without an engineer does not get used.
| Level | Builds | What it asks of you |
|---|---|---|
| Easy | 13 | A non-technical operator can run it the same day. |
| Medium | 85 | Copy-paste fluency required. No coding, but real setup. |
| Technical | 9 | Needs a developer to install and wire up. |
Client engagement outcomes, reported separately
These come from individual client engagements, not from the library dataset, so they are listed apart from it. Each is a single result rather than an average across a client base.
3 months to 2 weeks
Due diligence cycle time on a live client engagement.
40 hours a month
Manual reconciliation work removed for one client.
80% auto-matched
Incoming checks matched automatically with dual-model verification, with the remainder routed to a human.
Method, and what this data cannot tell you
The dataset is the public build library on this site, counted on 2026-07-20. Each entry carries an origin tag, a domain category, and a difficulty rating, and every count above comes from those fields directly. The library is open, so the numbers can be recounted by anyone who wants to check them.
The honest limitation is selection. This measures what we chose to build, and we build where an existing tool was not already doing the job. That biases the sample toward custom work by construction. It is evidence about the gaps we keep hitting in finance workflows, not a census of what software exists.
Two smaller caveats. The sample is one firm's work, so it reflects the kind of client that comes to us, mostly finance teams between five and eighty people. And the origin tag is a judgment call at the time of building: a workflow with no usable tool in 2025 may have one now, which is an argument for rechecking rather than for assuming the gap is permanent.
Frequently asked questions
- What share of finance AI workflows have no off-the-shelf tool?
- In our own published library of 107 builds, 79% (85 builds) had to be built from scratch because no existing project did the job. Within finance and data specifically the figure is 94% (73 of 78). The remaining 21% started from an existing open-source project and were packaged into a business workflow.
- What is the dataset behind these numbers?
- The public build library on consultance.ai, counted on 2026-07-20. Each of the 107 entries is a runnable workflow with its own page, a difficulty rating, a domain category, and a link to its upstream repository where one exists. The counts in this guide come from that dataset directly, so anyone can recount them.
- Does this prove no off-the-shelf finance AI tools exist?
- No, and it should not be read that way. It measures which workflows we chose to publish builds for, and we tend to publish where an existing tool did not already solve the problem. Read it as evidence about the gaps we keep hitting in finance work, not as a market-wide census of available software.
- Why is the off-the-shelf gap worse in finance than elsewhere?
- Three reasons show up repeatedly in the builds. The inputs arrive unstructured, as PDFs, statements, and email threads rather than clean API data. The output has to cite its source, because a number nobody can trace is unusable in a diligence or close process. And the workflow is usually shaped by one firm's conventions, so a generic tool solves half the problem and leaves the expensive half.
- How many of the builds are usable without a developer?
- 98 of 107 are rated Easy or Medium, meaning copy-paste fluency rather than coding. 9 are rated Technical and need a developer to install. Separately, 27 ship with a live interactive tool that runs in a browser.
- Are the client outcome numbers from the same dataset?
- No. They come from individual client engagements and are reported separately from the library counts on purpose. They are single-engagement results, not averages across a client base, and the sample is small enough that they should be read as examples rather than benchmarks.
Check the dataset yourself
All 107 builds are free and runnable. Recount the numbers, or take one and run it against your own workflow this week.