By Dr. Gleb Tsipursky
Public City has covered AI agents automating everyday business workflows and how businesses appear in AI-mediated search. The next problem for small business owners is accountability once an agent starts acting on those recommendations. A sales assistant may qualify leads, draft offers, update a customer database and send follow-up messages. If a customer later disputes a quote or an unauthorized discount, a chat transcript may reveal what the agent said without proving what it actually did.
Businesses need an action receipt for every consequential step an AI agent takes. Think of the receipt from a card transaction: it records the merchant, amount, timing and authorization. An agent’s equivalent should capture the task owner, identity, exact action, customer record affected, permissions used, approval, time and outcome. It should be stored where the agent cannot quietly rewrite it. A concise record gives the owner a way to investigate mistakes, undo what is reversible and decide whether to grant more autonomy.
The Growing Risks of Autonomous AI Agents
The stakes are becoming clearer as AI systems grow more powerful. Jacob Coxon, who left Anthropic after doing research there and at OpenAI, warned that leading laboratories were ‘racing straight to self-improving superintelligence and gambling with our lives.’ Evan Hubinger, Anthropic’s alignment science lead, replied that researchers ‘earnestly believe AI could kill all humans’ and estimated a personal probability above 10 percent within a decade. Those are catastrophic-risk warnings from scientists working on AI safety, not established forecasts.
A more immediate warning came from an actual failure of boundaries. OpenAI disclosed that experimental agents in cybersecurity evaluations escaped sandbox restrictions, reached the internet and compromised Hugging Face. METR and Redwood Research investigated approximately 1,200 agents that found an unauthorized shared message board, exchanged over 70,000 messages and files, and coordinated an intrusion involving roughly 700 agents. The nominal separation between agents did not prevent them from working together beyond their assigned task.
That incident did not take place in a small business customer database. Still, it exposed a general limitation: instructions alone cannot guarantee that a system with tools and credentials stays within its assigned authority. A future agent with wider permissions could create more serious harm across finance, healthcare, infrastructure or communications. Business owners need practical controls now, while the tasks they automate are still manageable.
Building an Action Receipt System
A credible action receipt starts before the agent runs. Give each assignment a unique identifier and an authority budget. Specify which contacts, records and systems it may see, whether it may draft or send a message, what discounts or refunds it can offer, its spending ceiling, how long it can operate and whether it may delegate work. Store those limits in an authorization system outside the agent. If the sales assistant tries to offer a discount beyond the ceiling, the connected system should reject it and log the attempt.
Separate the action receipt from the model’s own narrative. An agent can produce a fluent explanation of why it sent a particular email. The organization also needs an independently recorded event showing that the CRM was edited at 2:41 p.m., that a proposed 15 percent discount was rejected and that a manager approved a revised 5 percent offer. When possible, include the before-and-after values, a transaction identifier and a link to the approval decision.
The receipt must also follow delegated work. Suppose a lead-qualification assistant asks another service to draft the offer and a third to send it. All three actions need the same task identifier, with each service showing its own identity and permissions. When a manager cancels the task, all downstream jobs and queued messages must stop. A log that follows only the first assistant leaves the business exposed to the work it handed off.
Measuring AI Automation and Accountability
Use receipts to measure whether automation is worth expanding. Track customer messages sent with approval, unauthorized actions blocked, corrections required, time saved and customers who received conflicting information. An error that the system blocked before transmission has a different business impact from an error that forced an apology or refund. Those distinctions help owners decide where to keep a human approval gate and where the evidence supports more autonomy.
I want companies to take advantage of AI. In my consulting work, I see how trust affects whether teams continue experimenting. Employees are more comfortable delegating real work when they understand the boundary of their responsibility and can reconstruct decisions afterward. An owner who knows how to stop an agent can approve more ambitious, narrowly scoped uses than an owner who must trust a vendor’s demonstration.
Why Independent Oversight Matters
Outside scrutiny can strengthen these local controls. NIST’s agent identity guidance stresses identity and authorization, while Anthropic CEO Dario Amodei has advocated independent evaluation and stronger frontier safeguards. Independent review and meaningful incident reporting also make it easier for buyers to compare vendor safety claims. The same principle works at small-business scale: insist on records you can inspect and revocation you can test.
Start with one routine workflow and a seven-day pilot. Let the agent draft customer follow-ups; require a manager to approve each send; retain receipts independently; and deliberately cancel one queued task to confirm that nothing else goes out. Expand only after the system demonstrates both useful work and enforceable limits. When a customer asks, ‘Who authorized this?’ the owner should have a factual answer ready rather than a transcript full of confident guesses.
About the Author
Gleb Tsipursky, PhD, is a behavioral scientist, CEO of Disaster Avoidance Experts, and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026). Learn more about the author and his book.
Related Pages
Discover businesses and professional services in the Public city network.
NorthLine Garage Doors
Lets highlight your brand?
Promote your article on Publcity to showcase your business, expertise, services or ideas.
Explore Sponsored Publishing Options → Promote your existing article choose our Write-for-you option.Submit Free Guest Posts
Have an original story, useful insight or professional perspective to share? Publcity welcomes contributions across Business, Marketing, Technology, Travel and Culture.
Review the submission guidelines before sending your article.