Before you automate a role, process, or workflow, define what has to change in the business for the investment to be worthwhile.
I started with a pretty simple question:
How much of a growing consulting company can I automate before I actually need to hire people?
There's no shortage of solopreneur hype telling us that AI will let one person run a company that used to require ten people. Maybe. But if you're actually running a small business, the economics of that decision are far more important than the question itself.
As a company grows, so does the work around the work. Business development. Research. Recruiting. Finance. Administration. Coordination. Reporting. Proposal support. Project management. Eventually the instinct is to hire. And every one of those hires increases the cost structure of the business.
So I started experimenting with something.
Start somewhere
The first place I applied this to was business development. Opportunity discovery can consume an enormous amount of time. Searching contracts. Looking at companies. Monitoring hiring signals. Researching incumbents. Looking for potential partners. Comparing opportunities against past experience. Figuring out whether something is even worth another fifteen minutes of attention.
I had AI help automate much of that. It gathers information, organizes it, compares opportunities against my past work and company capabilities, filters obvious poor fits, and surfaces the things that appear most relevant.
That part works surprisingly well. But the experiment also exposed the limits pretty quickly. AI is much better at gathering and processing the information than deciding what I should actually do with it.
The system may tell me an opportunity is an excellent fit. I may look at it and immediately know it isn't. Maybe the customer is inaccessible. Maybe the incumbent is too entrenched. Maybe the technical match looks good on paper but isn't strategically useful. Maybe another opportunity deserves the time more. Maybe the opportunity itself is mediocre but the company behind it represents a valuable relationship.
That's judgment.
This led me to an important question to ask before automating almost anything:
Which part of this role exists because information is expensive to collect and process?
If most of the role consists of finding, reading, categorizing, comparing, summarizing and moving information around, AI may be able to remove a lot of the workload. It's the first actual test of whether AI makes sense.
But if the value of the role comes from judgment, relationships or actually executing a decision, eliminating the information burden doesn't eliminate the need for a person.
It could, however, change the type of person you need. Instead of hiring someone full-time to spend 30 hours gathering information and 10 hours exercising judgment, perhaps you automate the 30 hours and buy 10 hours of much better judgment.
That was my original theory: automate information, buy fractional human judgment.
On the surface, that seemed broadly applicable. But it doesn't actually apply to all business types.
Lower overhead is not always the right objective
For most small businesses, reducing overhead sounds unquestionably good. Government contractors, on the other hand, are more complicated.
Indirect costs such as G&A contribute to the contractor's rate structure. Some of those costs may ultimately be recovered through contracts depending on contract type, allowability, allocation and accounting treatment.
So simply saying "automation saves overhead" doesn't solve for everyone. A GovCon doesn't necessarily want the smallest possible indirect organization. It wants enough infrastructure to support growth while keeping its rates economically competitive.
A $150,000 business-development hire may increase indirect spending. But if that person helps support several million dollars of profitable new contract activity, the investment may be excellent. Conversely, a $25,000 AI automation may technically reduce labor while producing almost no meaningful financial benefit.
That changed the question for me again. It became:
What measurable economic result does the automation create?
And that distinction seems applicable far beyond GovCon.
Hours saved are not business value
We've all seen some version of the corporate productivity story: "This AI initiative saved employees 5,000 hours."
That sounds impressive. But there's a follow-on question that almost never gets resolved: what happened because those 5,000 hours were saved?
Did the company avoid another hire? Did revenue increase? Did throughput increase? Did cycle time fall? Did customers get served faster? Did margin improve? Did errors or risk decrease? Did working capital improve? Were those employees able to take on additional productive work?
If none of those things happened, what exactly was the point of the investment?
This starts to look like a new type of productivity theater. We already understand the old version. Looking busy is not the same thing as being productive. A person can spend all day moving between meetings, spreadsheets, email and reports and appear extremely busy without producing much economic value. AI gives us the ability to digitally repeat this same illusion.
A workflow can run 10,000 tasks. An agent can produce 2,000 summaries. A dashboard can report 4,000 hours "saved." None of that proves the business became more valuable.
Automated activity is still just activity.
If it doesn't have a positive business impact, the automation isn't earning its keep. And since labor is the most frequently cited savings for AI, here is one way to think about it:
Realized Labor Value = Hours Saved × Labor Cost × Capture RateThe important term is capture rate: what percentage of the theoretical savings actually becomes economic value.
If automation saves 500 hours but payroll, headcount, throughput and revenue remain unchanged, the labor capture rate may effectively be zero. Savings become real when the freed capacity does one of three things:
- eliminates or avoids paid labor,
- delays a future hire,
- or gets redeployed into measurable productive output.
So after claiming "500 hours saved," the next question should always be: what happened to those 500 hours?
If there is no measurable answer, don't assign them a dollar value. That sounds obvious when written down. Yet much of the current discussion around AI productivity focuses on labor hours saved.
Sometimes the original value hypothesis is wrong
I recently encountered this on an internal AI-assisted decision-support system build.
The primary user was asked to rate the value of the product from 1 to 10. The answer was an 8. The reasoning was essentially that, for where the product was in its lifecycle, it compared very favorably with other products in its category.
That sounds encouraging. But it doesn't tell the owner whether the investment is working. For a system like this, the owner's ROI was measured by how quickly his staff could take immediate action based on the information provided. So the obvious question was: how much time was it actually saving?
The answer was essentially none.
In terms of ROI, it seemed like it was missing the mark. But to the staff, the actual benefit was better insight. The system surfaced relationships and possibilities that they might otherwise have missed. Which increased the volume of decisions and actions taken downstream. Does this make it valuable? Not because it allowed faster response times, but it could be valuable in making better decisions.
This creates a causal chain that has to be proven. Does better insight lead to greater confidence, which influences faster or better actions, which generates improved outcomes, with economic value?
If the better insight never leads to faster or more accurate actions, the value stops there. If it improves actions but the outcome doesn't improve, the value stops there. Eventually something has to reach a measurable business result.
That realization led me toward a framework I'm now using to think about automation investments.
Make automation EARN continued investment
Before committing serious money to an automation initiative, I think four things should be clear.
(E)stablish the baseline
What measurable condition exists today? Not "this process is inefficient." What is inefficient? How many people does it consume? What does it cost? What is the error rate? What is the conversion rate? What is the cycle time?
Without a baseline, improvement is mostly opinion.
(A)rticulate the connection to value
How does the automation actually cause economic value? "AI improves research" is not enough. Maybe the chain is:
better research → better prioritization → fewer wasted pursuits → more capture resources concentrated on viable opportunities → improved win economics.
Or:
better insight → more confident action → faster response time → more successful transactions.
Whatever the chain is, write it down. If you cannot explain how the automation reaches a business outcome, you probably don't yet have an ROI hypothesis.
(R)ealize the value
Can the theoretical benefit actually be captured? As I've pointed out, saving labor does not automatically save money. Reducing cycle time does not automatically create revenue. Increasing capacity matters only if that capacity is needed. Improving insight matters only if someone acts differently because of it.
This is the step where a lot of theoretical AI ROI disappears.
(N)ame the hurdle
What exact outcome is required before the investment becomes worthwhile? This is the part I think organizations often skip.
If an automation costs $100,000 and produces $105,000 of theoretical benefit under perfect assumptions, that is not an attractive investment. Implementation estimates will be wrong. Adoption will be imperfect. Maintenance will cost something. People will work around the system. Edge cases will require manual intervention. Some theoretical savings will never be captured.
This is why automation needs margin for error.
Automation needs an X factor
Similar to switching costs when changing a vendor, automation requires some X factor to justify switching to an automated solution. Not a universal threshold for every automation, but some hurdle that must be achieved before it gets a green light.
As one example, I might want a low-risk, easily reversible internal automation to demonstrate something like 2–3 times its cost in expected realized benefit. A cross-functional automation requiring substantial behavior change might need 3–5 times. A revenue-critical or customer-facing system may need even more. Replacing an existing system that already works should require an improvement large enough to justify switching costs, retraining, migration risk and disruption.
Those numbers are not accounting standards. They are a way to force the question:
How much better does this actually have to be before we should bother?
Once you have that number, you can work backward.
Suppose an automation will cost $80,000 and management establishes a 3x hurdle. It now needs to create roughly $240,000 in realized economic benefit over the evaluation period. If each additional successful transaction contributes $20,000 of gross profit, the system needs to create approximately 12 incremental transactions.
Now we have something testable. What intermediate metric has to change to create 12 more transactions? How much faster must we act? How much more accurate must the system become? How often must its recommendations produce decisions that would not otherwise have occurred?
Suddenly the automation project has acceptance criteria tied to economics rather than features. And we can test how feasible the numbers are before any code is even written.
This changes how I think about staffing too
This brings me back to the original solopreneur question.
The answer probably isn't "automate everything so you never have to hire." And it probably isn't "AI will replace the back office."
The more interesting angle is how AI is integrated into jobs and roles.
Historically, a role may have required 30 hours of information work plus 10 hours of high-value judgment. Those activities were bundled together because information processing required humans. But AI now allows businesses to separate them.
Automate the information layer. Then decide how much judgment is actually required. Perhaps the answer is still a full-time employee. Perhaps it's a fractional CFO, recruiter, capture executive, architect or analyst. Perhaps the automation itself doesn't make economic sense.
The point is that the decision can now be made from the economics rather than from the historical assumption that every growing workload eventually requires another full-time person.
For a GovCon, that becomes especially interesting because indirect-cost growth can affect rate competitiveness. The question is not simply whether automation reduces G&A. It is:
How much additional profitable business can the indirect organization support for every additional dollar invested?
If automation allows the same indirect team to support significantly more contract volume, that may matter more than whether it eliminates a particular salary. If fractional expertise allows the company to purchase expensive judgment only when needed rather than carry it permanently, that may improve cost structure flexibility.
But again, it only matters if the economics actually improve.
Simplify before you automate
There is one question that belongs before all of this:
Should this process exist in its current form at all?
A twelve-step admin process does not become innovative because AI now performs all twelve steps. Sometimes the best automation is deleting six steps.
So the sequence becomes:
Simplify → EARN → verify
AI should have to EARN its keep
I'm still experimenting with all of this. Business development is one use case. Financial planning is another. I'm looking at research, project workflows and other internal processes the same way.
And I'm becoming much more skeptical of both extremes of the AI conversation. I don't think the right conclusion is "AI will let one person run an unlimited company." I also don't think "AI productivity is all hype." I think the useful middle ground is much more realistic:
AI should be evaluated like any other investment.
What metric is it supposed to change? What is the baseline? What mechanism creates the value? Can that value actually be captured? How much improvement is required to justify the investment? And what happens if our assumptions are wrong?
If we can't answer those questions, we may still have a worthwhile experiment, which is valuable — but it should be properly funded and evaluated as an experiment.
If automation doesn't change a business outcome, it isn't transformation. It's probably just automation theater. And increasingly, I think that is the distinction businesses need to make before they decide where AI actually belongs.