Asher's Blog

Founders usually define competitors as companies that sell a similar product. If you are building payroll software, you list other payroll companies. If you are building accounting software, you study other accounting platforms. If you are building an AI assistant, you compare it with other AI assistants.

But in many markets, especially when selling software to businesses, your most important competitor is not another company.

It is the customer saying, “I’ll do it manually.” The customer already has a spreadsheet. Someone in the office knows how to complete the form. The finance manager can initiate the transfer from a banking application. Employees can send their information through WhatsApp. An administrator can copy the figures from one system into another. The method may be slow, fragmented and vulnerable to mistakes, but it is familiar, and familiarity is a formidable competitor.

This is not peculiar to Nigeria. I have visited businesses in the United States and found important processes still being completed through spreadsheets, emails, phone calls and people moving information between systems. Technology may be more available, but the work does not automatically become automated merely because software exists.

Many businesses are held together by human beings performing invisible acts of coordination.

That is why one of the most useful ways to find a valuable business idea is to look for places where people are creating significant value through repetitive manual effort. Observe the tasks that require somebody to copy, check, remind, reconcile, interpret, approve and follow up. Whenever an important outcome depends on a person repeatedly carrying information between disconnected systems, there may be an opportunity to build.

The first generation of software helped people organize that work. The emerging generation of AI agents may increasingly perform parts of the work for them.

Manual work is often the incumbent

When a founder says there is no competitor, I become suspicious. The problem is already being handled somehow, even if the method is inconvenient.

Perhaps a business owner calculates payroll in Excel. Perhaps an accountant maintains several versions of the same report. Perhaps someone signs into three banking applications every morning to review balances. Perhaps an HR manager checks employee credentials through calls and emails. Perhaps a customer service officer reads a complaint, looks up the transaction in another system, sends a message to operations and returns later with an answer.

Those methods are the incumbent product. They have no sales team and may not appear in market reports, but they benefit from something every challenger should respect: the customer already knows how to use them. The spreadsheet may be ugly, but it contains years of exceptions and institutional knowledge. The employee who manages the process knows whom to call when something goes wrong. The manual method has been adapted gradually to the organization’s particular reality.

A new product cannot win merely by being more modern. It must produce enough additional value to justify the cost and discomfort of change.

This is where many founders misread the market. We see a process that looks inefficient and assume the customer is waiting for software. But inefficiency alone does not create demand. The customer must believe that changing the process will save enough time, reduce enough risk, increase enough revenue or create enough control to make the transition worthwhile.

The real comparison is rarely your feature list against another company’s feature list. It is your product against the customer’s existing combination of people, spreadsheets, messages, memory and habit.

Look for concentrated manual value

Not every manual task should become a company. Some tasks are too rare. Some are inexpensive to perform. Some require judgment that customers do not want to delegate. Others occur inside markets too small to support a meaningful business. The opportunity becomes more interesting when several conditions appear together.

The work is frequent; it consumes meaningful time. Mistakes are expensive. Several systems or people are involved; the process generates useful data. The customer already pays someone to perform it. The outcome affects revenue, compliance, cash, safety or customer experience. And the process can be made substantially better without removing judgment where judgment is still necessary.

These are signs that manual effort is carrying a large amount of value.

Payroll is a good example; it is not merely data entry. Someone must collect employee information, account for new hires and exits, calculate variable payments and deductions, confirm funding, initiate payments, produce payslips, reconcile transactions and manage statutory obligations. In many businesses, several people and systems participate in the process.

The opportunity is not simply to digitize the final payslip. It is to understand the entire sequence of work and decide which parts can be eliminated, automated, combined or made more reliable.

The same pattern exists in procurement, logistics, compliance, customer support, healthcare administration, insurance claims, property management and financial reconciliation. Wherever skilled people spend much of their day moving information rather than applying their judgment, there is room to reconsider the workflow.

Manual work is also the prototype

There is another side to this argument that founders should not miss. Manual work is not always the enemy. Sometimes it is the prototype.

Before automating a process, it is often useful to perform it manually and observe the decisions hidden inside it. What information does the employee check? Which exceptions occur repeatedly? When does the person ignore the official process because reality does not fit it? What judgment is being applied that nobody has documented?

If you automate too early, you can build a fast version of the wrong process.

Bill Gates made a useful observation in Business @ the Speed of Thought: automation applied to an efficient operation magnifies efficiency, while automation applied to an inefficient operation magnifies the inefficiency. The principle has become even more important in the age of AI. An agent can repeat a bad decision more quickly, more cheaply and at a much larger scale than a person.

The founder’s task is therefore not to look at a manual process and say, “Let us automate everything.” The task is to understand what creates value, remove what should not exist, preserve necessary judgment and automate the parts that are repetitive, predictable and verifiable.

The manual process teaches you what the real product should be. This is related to the advice Paul Graham gives startups to do things that do not scale in the beginning. Early manual work allows founders to get close to the customer and learn the shape of the problem. But what is appropriate as a learning mechanism can become a constraint when the company grows. The founder must know when the manual process has produced enough understanding to be converted into a system.

Do it manually to learn. Do not continue doing it manually because you failed to learn.

From tools to workers

Traditional software usually waits for instructions. The user opens the application, navigates through the interface, enters information, selects an action and confirms it. Software makes the work faster, but the human being continues to coordinate the sequence.

AI agents introduce a different possibility. Instead of helping with one step, an agent can be given an objective, gather relevant context, use approved tools and execute a sequence of actions. The interface begins to move from clicking through menus towards expressing intent.

This is why agents matter; an AI chatbot gives an answer. An agent is expected to do work. It may search records, compare information, prepare a document, update a system, initiate a workflow, monitor the result and return when a decision or approval is required.

The Anthropic Economic Index offers a useful indication of this transition. Its early analysis of millions of Claude conversations found that AI was being used both to augment people and to automate work. In the first dataset, augmentation was more common, but later reports observed a shift towards a greater share of automated interactions. The precise percentages will continue to change, and data from one platform cannot describe the entire economy, but the direction is important: people are increasingly asking AI not only to help them think, but to complete tasks with less continuous involvement.

Microsoft’s 2025 Work Trend Index described a similar expectation among business leaders, reporting that a large majority expected agents to become meaningfully integrated into their AI strategies within the following 12 to 18 months. These are company-sponsored studies and should not be treated as neutral forecasts of exactly what will happen. Still, they reflect what many product builders can already observe: the unit of software is beginning to move from the screen towards the delegated task.

Your assistant should understand the instruction

Imagine that I want to send ₦50,000 to my wife. Today, I may open my Zenith Bank application, sign in, choose the beneficiary, enter the amount, verify the details and authorize the transfer. The banking application has digitized what once required visiting a branch, but I am still performing and coordinating every step.

In a more agentic system, I should be able to tell a trusted personal assistant: “Send ₦50,000 to my wife from my Zenith account.”

But the truly useful assistant should do more than obey the literal instruction. It should understand the financial context in which the instruction is being made. It may tell me that completing the transfer would leave ₦80,000 in that account, remind me of a scheduled payment, show that I have more liquidity in another approved account or explain that using a particular credit line would create an avoidable cost.

It might say: “You asked me to send ₦50,000 from Zenith. That would leave ₦80,000, while a ₦100,000 payment is due tomorrow. You have enough available in Access Bank. Would you prefer that account?”

That is more useful than blind obedience; it combines instruction, memory, context and judgment support.

However, there is an important boundary. An agent that can discuss money is different from an agent authorized to move money. The second requires identity verification, explicit permissions, transaction limits, secure credentials, audit trails, fraud detection, revocation controls and clear responsibility when something goes wrong.

For some actions, the appropriate design may be for the agent to prepare the transaction and ask for final confirmation. For smaller, predictable payments to approved beneficiaries, the user may grant standing authority within defined limits. Large, unusual or irreversible transactions should require stronger human approval. The goal is not maximum autonomy; the goal is appropriate autonomy.

The future assistant should challenge you

The most useful personal assistant is not one that says yes to everything. It is one that understands your objectives well enough to notice conflict.

If I have set a savings target, the assistant should understand it. If I have several accounts and wallets, it should provide a consolidated view with my permission. If a transaction would create a cash shortage, it should alert me. If an investment is inconsistent with my risk preference, it should explain that. If a credit line is available, it should distinguish between using credit strategically and paying unnecessary interest merely because borrowing is convenient.

This is the movement from command execution to decision support. But it must be designed carefully. Financial decisions contain context an AI system may not know. The assistant may have incomplete information, misunderstand an instruction or optimize for a stated goal while ignoring an unstated priority. It should therefore communicate uncertainty, disclose assumptions and ask for clarification when the cost of being wrong is high.

An assistant that acts confidently with incomplete context can create more work than the manual process it was meant to replace. Trust will depend on whether the agent knows when not to act.

Find work by watching work

Founders looking for AI opportunities should spend less time brainstorming abstract applications and more time observing people at work.

Sit beside the payroll officer at the end of the month. Watch the accountant reconcile transactions. Follow the customer service representative through a difficult complaint. Observe how a logistics team responds when a delivery fails. Ask the compliance officer what information is repeatedly requested and why. Look at the tabs open on their computers, the spreadsheets they maintain and the messages they send to move a task forward.

Do not ask only, “What do you want AI to do?” Most people describe the future using the vocabulary of their current process. Ask them to show you what they did yesterday. Ask what took the longest, what required another person, what they had to check twice and what they were afraid of getting wrong.

The best automation opportunities are often visible in the handoffs. One person exports a file. Another cleans it; someone sends it for approval. A manager asks for an explanation; the file is uploaded elsewhere. A transaction fails. Someone returns to the first spreadsheet to correct the record. Each handoff introduces delay, ambiguity and the possibility of error.

An agent may eventually coordinate several of these steps, but only after the builder understands why each one exists.

Do not automate the human out of the relationship

There is a temptation to treat every human interaction as inefficiency. That would be a mistake.

Some human involvement exists because the customer needs empathy, accountability, negotiation or judgment. A bereaved insurance customer does not necessarily want to argue with an agent about a claim. An employee disputing a salary calculation may need a person capable of understanding an unusual circumstance. A founder facing a consequential financial decision may want counsel, not merely a mathematically optimized recommendation.

The purpose of AI is not to eliminate people from every process. It is to remove work that prevents people from contributing what is distinctively valuable.

If an agent can collect the information, check the records, prepare the analysis and identify anomalies, the human professional can spend more time exercising judgment, communicating and taking responsibility. In many settings, the winning design will be neither fully manual nor fully autonomous. It will be a thoughtful combination of machine speed and human accountability.

This is why the National Institute of Standards and Technology’s AI Risk Management Framework emphasizes governance, measurement and the definition of roles in human-AI configurations. As systems gain the ability to act, organizations must be explicit about who has authority, who monitors performance and who remains accountable for the outcome.

An AI agent may execute the work; it cannot absorb the company’s responsibility.

The hidden cost of manual work

Manual processes often appear cheap because their costs are spread across salaries and time.

The company may say, “We already have someone who does that,” without calculating how many hours the task consumes, how often it delays another process, how much rework errors create or what opportunities the employee cannot pursue because they are occupied with administration.

There is also the cost of key-person dependence. When only one employee understands the spreadsheet or knows the sequence of calls required to complete a process, the company does not truly own that process. It is renting institutional memory from one person.

Automation can make the process visible, measurable and repeatable. It can preserve knowledge, create audit trails and allow the company to improve performance over time.

But the business case must be honest. If the manual process takes ten minutes a month, an expensive AI implementation may not create meaningful value. If the work changes constantly and errors would be catastrophic, full automation may be premature. A founder should calculate the economic value of time saved, errors prevented, throughput increased and risk reduced.

AI is not automatically valuable because it is AI. It is valuable when it changes the economics or quality of an important outcome.

The opportunity for African founders

This way of thinking is particularly important in Africa because so many businesses operate through fragmented processes.

The absence of integrated systems means people perform a great deal of coordination themselves. Data is copied between applications. Payments are checked manually. Compliance information is gathered from several sources. Business owners depend on employees who understand how to navigate inconsistent infrastructure and changing requirements.

At first, this can make software adoption harder. The product must fit a reality filled with exceptions, limited data and informal workarounds. But it also means there is enormous value available to companies that understand the workflow deeply enough to simplify it.

We should not merely import AI products designed around American assumptions. We should study the manual work being done in Lagos, Nairobi, Accra and other African cities. We should ask why the workaround exists, what institutional gap it is covering and how an intelligent system could make the process more reliable without requiring perfect infrastructure.

The opportunity may not look like building the largest foundational AI model. It may be applying intelligence to payroll, tax compliance, logistics, healthcare administration, credit assessment, agriculture, procurement or government services.

The value will come from understanding the local work better than anyone else.

Never recommend the old way by default

Recently, I needed to complete a particular task. An AI system had initially suggested a traditional method, and I paid for it. I later discovered an AI-powered tool capable of doing the work more directly. The difference was significant enough that I requested a refund from the first provider.

I told the assistant not to recommend the traditional method to me again when a credible AI alternative exists.

That experience captured how quickly customer expectations are changing. People who discover that a task can be completed in minutes will become less willing to accept a process that takes days. Once an intelligent tool performs work that previously required repeated interaction with a service provider, the old method begins to feel unnecessarily expensive.

This does not mean the AI option is always correct. Tools must still be assessed for accuracy, privacy, security and suitability. But every business should ask an uncomfortable question: if an AI agent can perform the valuable part of our service faster and more cheaply, what exactly will customers continue paying us for?

The answer may be proprietary data, trusted execution, regulatory permission, human judgment, distribution, accountability or a superior end-to-end experience. If the answer is merely that customers have not discovered the alternative yet, the company is vulnerable.

Build for outcomes, not interfaces

The deepest change created by AI agents may be that customers will care less about using individual applications.

Today, every company asks the customer to learn its interface. Tomorrow, customers may increasingly express an objective to an assistant and allow that assistant to coordinate approved services on their behalf. The assistant may select the account, prepare the transfer, compare providers, retrieve documents, fill forms and monitor completion.

This means companies should think beyond owning the screen. They must build products that can be securely understood and used by other software through reliable APIs, clear permissions and verifiable outcomes.

The company that insists every customer must manually navigate its application may be competing with a future in which the customer does not want to open the application at all.

The winning product will not necessarily be the one with the most impressive dashboard. It may be the one that completes the customer’s objective most dependably, whether the instruction comes from a person or an authorized agent.

Start with the sentence “I have to…”

If you are looking for an idea, listen for people saying:

“I have to copy this every week.” “I have to check all these accounts.”

“I have to remind them before anything happens.” “I have to download it, edit it and upload it again.”

“I have to call someone who knows how it works.” “I have to verify it manually because the systems do not talk to each other.”

These sentences are maps of unfinished technology. Then ask four questions. Is the outcome important? Is the work repeated? Can the system observe enough context to perform it reliably? Can the action be verified or reversed when something goes wrong?

If the answers are strong, there may be a real opportunity. Your most important competitor may not have a brand, a website or a fundraising announcement. It may be an employee with a spreadsheet, a WhatsApp group, three browser tabs and a process that everybody complains about but nobody has replaced.

Respect that competitor. Learn from it. Understand the exceptions it has absorbed and the trust it has earned.

Then build something meaningfully better. The future of work will not arrive merely because companies add AI to their descriptions. It will arrive as intelligent systems gradually learn to perform valuable sequences of work, while people retain appropriate authority, judgment and accountability.

The opportunity is enormous, but the starting point is ordinary: watch what people are still doing by hand.

That is where the work is. And very often, that is where the company is waiting to be built. —

References and further reading

• Paul Graham, “Do Things That Don’t Scale”.
• Bill Gates, Business @ the Speed of Thought: Succeeding in the Digital Economy.
• Michael Hammer and James Champy, Reengineering the Corporation.
• Anthropic, The Anthropic Economic Index and 2026 research on automation and augmentation.
• Microsoft, 2025 Work Trend Index: The Frontier Firm Is Born.
• National Institute of Standards and Technology, AI Risk Management Framework.
• United States Department of the Treasury, Artificial Intelligence in Financial Services.