Guide · See
How to find where AI belongs: AI discovery vs. consultants vs. process mining
Workshops, consultants, process mining, task mining and AI discovery compared: what each sees and misses, and how to rank AI use cases by value and readiness.
The short answer
Find where AI belongs by mapping how the work really runs, then ranking each opportunity by value and readiness. Workshops and consultant assessments capture what people report, process mining shows what systems record, and task mining shows what happens on desktops. AI discovery combines people, systems and documents, and when it runs continuously for every team, the map stays current as the work changes.
Key takeaways
- Each approach sees a different slice of the work. The most complete picture combines people, systems and documents.
- Rank what you find by value (how often × how long × how many people × the cost of their time) and by readiness (judgment, data access and risk).
- Don’t start from tools, demos or the documented process, and don’t rely on one person. Check what people say against how the work actually runs.
What are the main ways to find where AI belongs?
Companies look for AI opportunities in six main ways: workshops, surveys and idea portals; consultant-led assessments; process mining; task mining; discovery led by forward-deployed engineers (FDEs); and continuous AI discovery. Each one sees a different part of the work.
That matters because work leaves evidence in four places. Systems of record show what happened and how often. People know the reasons behind each step and how exceptions get handled. Documents say how the work is supposed to run. And much of the effort happens between apps, in email, spreadsheets and chat, where no system logs it.
| Approach | Systemswhat was logged | Peoplewhy, and unwritten rules | Docshow it should run | Between appsemail, sheets, chat |
|---|---|---|---|---|
| Workshops, surveys and idea portals | Not seen | Partly | Not seen | Partly |
| Consultant-led assessment | Partly | Partly | Main source | Partly |
| Process mining | Main source | Not seen | Not seen | Not seen |
| Task mining | Not seen | Not seen | Not seen | Main source |
| FDE-style discoveryOne team at a time | Main source | Main source | Main source | Main source |
| Continuous AI discoveryEvery team, kept current | Main source | Main source | Main source | Partly |
- Main source
- Partly: a sample, or what people recall
- Not seen
No approach sees everything. What separates them is which slice they show, what they miss, and how long their picture stays true.
What do workshops, surveys and consultant assessments find?
Workshops, surveys and idea portals ask employees where AI could help. They are a common first step. OpenAI’s guide to finding use cases, for example, suggests hackathons, company-wide competitions and a shared spreadsheet or Slack channel to collect ideas, then an impact-and-effort grid to rank them.
They are fast and cheap, and they build energy: people learn what AI can do and own the ideas they bring. The limit is that you only hear what people think to mention. A workshop rarely surfaces how many hours a task takes across the company, how often its exceptions happen, or which steps were never written down. OpenAI also notes that most teams start with individual tasks, which are easier to picture than multi-step workflows.
Consultant-led assessments add structure and an outside view. A team interviews leaders and experts, reviews documents and data, and scores the options. PwC’s AI Readiness Assessment, for example, scores a company on more than 285 diagnostic questions across 12 domains, compares it with over 500 organizations, and ends in a 90-day, six-month and 12-month roadmap.
That is valuable when a board needs a strategy: where the company stands, how it compares with peers and what to do first. The trade-off is that an assessment captures one moment, mostly through what people report. It works at the level of the company or the function, and it starts to age once the engagement ends.
What do process mining and task mining show?
Process mining reads the event logs that business systems already keep, in tools such as an ERP or a CRM, and rebuilds how a process actually ran: every case, step, loop and wait. The IEEE Task Force on Process Mining’s manifesto describes the idea as discovering, monitoring and improving “real processes (i.e., not assumed processes)” from those logs.
Its strength is precision. It counts how often each path happens and how long each wait lasts, and its conformance checking shows where the real process departs from the intended one. On The Startup Ideas Podcast, Varick Agents described one engagement: a $5 billion software company’s process document listed seven steps to get a quote signed. Process mining on its CRM found 20 steps and seven loops, and 61% of requests looped back to the start.
It has two limits. Clean event data takes work to get: Wil van der Aalst of RWTH Aachen University and Celonis notes that companies keep running into data scattered across dozens of systems and tables, which makes extraction slow; newer object-centric methods aim to ease it. And process mining sees only what systems record. Work done in email, spreadsheets and chat leaves no event, and a log can’t say why someone made a choice.
Task mining covers part of that gap. Software on employees’ computers records their clicks, keystrokes and data entry, and sometimes screenshots, as IBM describes it. Celonis, which sells both kinds of tools, pairs it with process mining to catch the steps outside major systems, like checking an email or a spreadsheet.
That makes the copy-and-paste work between apps visible. But task mining records what people do without the reasons, and IBM notes that it can lose the larger context of the process. Because it records people’s screens, it also raises privacy questions; IBM advises that users approve these tools before they are switched on.
How does AI discovery combine people, systems and docs?
AI discovery is the work of finding where AI belongs: which workflows, which steps within them, and what each change is worth. Forward-deployed engineers do it by embedding with a team. They combine people, systems and documents, and they watch the work that happens between apps.
Varick Agents, an FDE firm, audits how the work really runs for at least four weeks before it touches a model, according to Daniel Kornum of Varick. Its FDE course lists what the audit draws on: interviews, emails, spreadsheets, SOPs, systems, approvals and exception paths. The result is one operating map: the workflow today, the workflow rebuilt around AI, the use case chosen, its boundaries, and its expected value in hours, cost and errors.
The sources keep each other honest. On the podcast, Varick’s CEO, Vasuman Moza, said that system data without interviews misses half the picture.
FDE-style discovery sees the most, but one team at a time. FDEs are scarce and expensive, and the map starts to go stale when they leave. See what forward-deployed engineers do.
Continuous AI discovery applies the same playbook to every team, all the time. It learns from people’s work on company devices, their email and calendar, and the logs of the AI agents they use, and it updates the map as the work changes.
The case for running it continuously is that processes change. The process mining manifesto lists concept drift, a process changing while it is being studied, among the field’s main challenges, and notes that few processes are in a steady state. Continuous discovery has limits too. It is newer, and it sees only what people share and what the company connects. Like any method that learns from people’s work, it needs their consent and clear rules.
AI discovery vs. consultants vs. process mining: how do they compare?
Here are the six approaches side by side: what they see and miss, how current they stay, who takes part, what they cost, and what you get.
| Approach | Sees | Misses | Stays current? | Who takes part | Time and cost | You get |
|---|---|---|---|---|---|---|
| Workshops, surveys and idea portals | Ideas and pain points people volunteer | Volumes, time spent, unwritten steps | No: a snapshot | Many employees and a facilitator | Days; mostly people’s time | A long list of ideas |
| Consultant-led assessment | Interviews, documents, sample data, peer benchmarks | Step-by-step detail; teams not interviewed | No: it ages after delivery | Leaders, selected experts, the consultants | Weeks; consulting fees | Ranked options and a roadmap |
| Process mining | Every case in the system logs | Work outside systems; the reasons why | Can, if data loads continuously | IT, data engineers and process owners | A slow data setup; software licenses | Process maps, loops, delays and deviations |
| Task mining | Clicks, keystrokes and apps on desktops | Reasons, judgment, off-screen work | Only while recording | Recorded employees, with their consent | A recording period; software licenses | Task maps and time per app |
| FDE-style discovery | People, systems and docs, in depth | Every team outside the engagement | While the engineers are embedded | Embedded engineers and the whole team | Weeks per workflow; senior engineers | An operating map, then an agent |
| Continuous AI discovery | Screen work, email, calendar and AI agent logs | Work outside the apps your company allows | Yes: it updates as the work changes | Set once, by your admin | Ongoing; runs in the background | A current map and ranked opportunities |
They work well together. A practical split:
- Workshops to build energy and collect leads. Check the ideas against the work before you rank them.
- A consultant assessment when the board needs a strategy, a peer benchmark and a sequenced roadmap.
- Process mining where the work runs through systems of record and you need exact volumes, loops and delays.
- Task mining to study one desktop-heavy task in detail, with the consent of the people recorded.
- FDE-style discovery to rebuild one high-stakes workflow now.
- Continuous discovery to keep a current map of every team, so you know where to look next.
How do you prioritize the AI opportunities you find?
Discovery usually turns up more opportunities than you can fund. Rank each one on two things: value and readiness.
Value is the time the work takes, priced: how often it happens × how long it takes × how many people do it × the cost of their time. Varick’s FDE course gives similar advice: prioritize long, high-volume workflows where the improvement is big enough to matter.
Readiness is how safely AI can take the work on today. Check three things:
- Judgment: does the work follow fixed rules, or does it need judgment you can check against past cases?
- Data access: can you reach the inputs, and do you have past cases with known right answers?
- Risk: what would a mistake cost, and can it be undone?
- Weekly revenue review$44k a year
- Ticket triage$166k a year
- Invoice matching$95k a year
- Renewal forecast$14k a year
- Deal desk approvals$25k a year
Show the data
| Workflow | Hours a month | People | Value a year | Readiness |
|---|---|---|---|---|
| Ticket triage | 210 | 8 | $166,320 | High: routing follows nine consistent rules; 3,700 tickets a month to test on |
| Invoice matching | 120 | 4 | $95,040 | High: 1,050 invoices a month to test on; a reviewer keeps the exceptions |
| Weekly revenue review | 56 | 6 | $44,352 | Medium: the data sits in five apps, and the figures conflict |
| Deal desk approvals | 32 | 4 | $25,344 | High: the same three checks in every region; a person approves |
| Renewal forecast | 18 | 3 | $14,256 | Low: renewal risk is a judgment call, checked only when renewals close |
In this example, ticket triage and invoice matching come first: they are worth the most and are ready now. The weekly revenue review is worth doing, but its data sits in five apps and its figures conflict, so its steps need sorting before anything is built. The renewal forecast stays with people for now: renewal risk is a judgment call, and you only learn whether it was right when renewals close.
For each workflow you pick, sort every step into the four buckets: delete it, run it as plain code, give it to an agent, or keep a person on it. Workflows rarely go to AI whole; usually only a few steps need an agent. Then prove the change on your own work, for example in shadow mode, before it goes live.
What mistakes do companies make when looking for AI use cases?
Five come up again and again:
- Starting from tools or demos. Kornum describes the usual path: pick the obvious workflow, scope it from the SOP, build it, demo it to the CFO and deploy. Then it collides with how the work actually runs.
- Trusting the documented process. Kornum calls the gap between the SOP and the real work the conformance gap. In typical Varick engagements it is 30% or more, and in exception-heavy work, such as accounts-payable exceptions, it is often over 70%. Build for the document and you automate about 70% of the volume and break on the rest.
- Asking one person. Even a confident expert describes their own version of the work, and Moza warns that one person can tell you the wrong story. Talk to everyone who touches the work, and check what they say against the systems.
- Ranking by excitement instead of value. OpenAI’s guide warns that impressive, complex use cases often slow teams down. Bain finds that many CEOs count pilots as progress, while the leaders concentrate on a few areas where AI changes the economics of the business.
- Treating discovery as a one-off. The work changes, and so does what AI can do. OpenAI suggests re-scoring use cases every quarter.
Questions people ask
What is an AI opportunity assessment?
A structured review, often run by consultants, that finds where AI could create value, scores the options and produces a roadmap. Most draw on interviews, workshops and documents, so they reflect what people report at one point in time. Pair one with evidence of how the work actually runs.
What is the difference between process mining and task mining?
Process mining rebuilds a process from the event logs of business systems such as an ERP or a CRM, so it shows the end-to-end flow: every case, loop and delay. Task mining records what people do on their computers, such as clicks, keystrokes and data entry, so it shows the work between the system steps. Neither one shows why people make the choices they do.
Is process mining enough to find AI use cases?
It is a strong start for work that runs through systems of record, because it measures volumes, loops and delays exactly. It can’t see work done outside those systems or the judgment behind each step, so combine it with interviews and documents before you decide where AI belongs.
How long does it take to find where AI belongs?
It depends on the approach. A workshop takes days. A consultant assessment or an FDE-style audit takes weeks for each scope; Varick Agents, for example, audits how a workflow really runs for at least four weeks before building. Continuous discovery never finishes: it keeps the map current as the work changes.
Who should be involved in finding AI opportunities?
The people who do the work, the people who own the systems, and a leader who can make decisions. Talk to more than one person for each workflow, because one person can tell you the wrong story, and check what people say against the systems.
How do you prioritize AI use cases?
Estimate value as how often the work happens × how long it takes × how many people do it × the cost of their time. Then check readiness: how much judgment it needs, whether you can reach the data and past cases, and what a mistake would cost. Start where both are high, and re-rank as the work changes.
Sources
- “Identifying and scaling AI use cases.” OpenAI, accessed October 3, 2026.
- “AI Readiness Assessment.” PwC, accessed October 3, 2026.
- IEEE Task Force on Process Mining (Wil van der Aalst et al.). “Process Mining Manifesto.” Business Process Management Workshops (BPM 2011), Springer, 2012.
- Greg Isenberg, with Vasuman Moza of Varick Agents. “Become a $1M/yr FDE (Full Course).” The Startup Ideas Podcast, October 1, 2026.
- Wil M.P. van der Aalst. “No AI Without PI! Object-Centric Process Mining as the Enabler for Generative, Predictive, and Prescriptive Artificial Intelligence.” arXiv, July 31, 2025.
- “What Is Task Mining?” IBM Think, accessed October 3, 2026.
- “What is Task Mining?” Celonis, accessed October 3, 2026.
- Daniel Kornum. “If AI is so great, why isn’t it working?” Varick Agents.
- “AI Forward Deployed Engineering: The blueprint to breaking into FDE in 30 days.” Varick Agents.
- “What Is an AI-Native Enterprise?” Bain & Company, August 2026.