AI is getting better at doing the work.
Small businesses still need to decide when that work is actually done.
> An invoice can be processed and still go to the wrong approver.
> A candidate can be marked ready while a required document is missing.
> A customer request can be answered without resolving the problem.
The automation ran but someone still has to check the result. That’s the problem we’re building Praxis around. Automate your business without losing judgment.
Here’s what that means for the people running the business.
1. The process usually lives in someone’s head
Ask a staffing agency how onboarding works and you’ll get the official version:
Collect the documents. Complete the checks. Update the record. Confirm the start date.
Then ask the person who actually runs it.
They know which documents need another look. Which client has an extra requirement. Who can approve an exception. What to do when the candidate is supposed to start tomorrow.
Those details keep the business working.
Before you can automate the process reliably, you need to make them explicit.
2. A small team can still have complicated rules
A five-person agency might handle different payment terms for every client.
A small staffing firm might track credentials, expiry dates and placement requirements across dozens of workers.
A clinic might receive invoices from multiple suppliers, for different locations, with different people responsible for approval.
Headcount tells you very little about how complicated the work is.
These teams need a practical way to describe how their business operates, including the exceptions they’ve learned through experience.
3. The happy path is the easy part
Consider an invoice workflow.
Read the email.
Extract the details.
Match the supplier.
Find the purchase order.
Prepare it for approval.
Now change one thing.
The supplier’s bank details are different.
Or the invoice looks like one you already paid.
Or the amount is correct, but the work hasn’t been accepted yet.
The next step depends on the business’s rules. It might require another document, an independent check or a decision from the person responsible.
That’s where the workflow needs to pause clearly, with enough context for someone to act.
4. Human approval needs to be a real decision
An “Approve” button is only useful if the person understands what they’re approving.
For an invoice, that could mean seeing:
The supplier and amount.
The purchase order it matches.
The checks that passed.
The discrepancy that remains.
The exact action that will follow approval.
If the reviewer has to reopen five tabs and reconstruct the whole case, much of the work has simply moved into the approval queue.
AI should help prepare the decision. The person should retain the authority to make it.
5. A green check isn’t enough evidence
A workflow can execute every step and still miss the requirement.
The connector ran, but the CRM record wasn’t saved.
The reminder went out, but the customer had already disputed the invoice.
The onboarding checklist was completed, but a credential had expired.
“Done” needs a definition.
For candidate onboarding, it might mean every required document is present, the relevant checks are recorded, and the authorised person has confirmed readiness.
For invoice processing, it might mean the details match, duplicates have been checked, and unresolved discrepancies are visible.
The business should be able to inspect that definition and change it when its requirements change.
6. Start with the chore everyone keeps chasing
You don’t have to begin by automating an entire department.
Start with the recurring task that keeps pulling someone away from their real work.
Chasing missing onboarding documents.
Preparing invoices for review.
Sorting incoming client requests.
Following up on unsigned forms.
Reconciling records between two systems.
Pick one.
Describe what arrives, what needs checking, who owns the decision and what should happen when something is missing.
Then test an awkward case alongside the straightforward one.
The incomplete record often teaches you more about the workflow than the perfect example.
7. The operator needs to be able to steer it
The person running the business knows why a particular client needs an extra check.
They know why a reminder should wait.
Why a placement needs review.
Why an apparently valid document doesn’t settle the matter.
They should be able to bring that knowledge into the workflow, inspect what’s proposed and correct its assumptions.
That’s the direction we’re exploring with Praxis.
A business describes the chore. We help shape the tools, steps, rules, checks and human approval points around it.
That structure is what we mean by a harness.
The goal is to give small teams a way to delegate repetitive work while keeping important decisions understandable and accountable.
Better AI makes more work possible.
Your business still needs a say in how that work gets done.
@i_mika_el Agreed. Approval should cover the exact ticket and its refund terms. If either changes before checkout, the agent should pause and ask again. Otherwise you can approve one thing and end up buying another.
Make one portfolio project survive a messy business workflow: missing documents, duplicate requests and an approval that never arrives.
What proves the work finished, and who owns the exception? That’s the harness we’re exploring at Praxis.
@vinayhjoshi@TravisLeeSwiftX No update from the warehouse’ is useful information. I’d carry it forward with a named owner and next check time, so it doesn’t disappear from next week’s report. Who usually owns chasing that missing update?
@quadrillionboss The DB check closes the loop. I’d also keep the booking version each attempt used: no duplicate doesn’t necessarily mean the latest details were applied. Would you pause for review if the booking changed between attempts?
@derecktafuma@developmintxyz We'd make the deadline visible to the owner first, with a reminder before it's missed. Showing it to the customer creates a promise the business needs to keep. Even a quick “Got your enquiry, I'll get back to you by 4” gives both sides a clear next step.
@BhagatDhanji That gives the invoice a clear source of truth. Does it keep a reference to the accepted quote version too? That would let someone checking a disputed line trace it back to what the client approved without comparing documents manually.
@ganesh2412@i_mika_el@Ingvar29007836@Michaelchoong5 And a new evaluation shouldn't trigger a new action by itself. If today's rules would allow a previously blocked payment, the operator still needs to confirm it's needed and hasn't already been handled. A changed verdict isn't fresh authorization.
@lindadoeswp That convenience matters. We'd want a fresh approval if the post changed, with the changes visible so you don't have to dig through wp-admin. We're exploring that balance with Praxis. Would love your take: https://t.co/NM24HFOzZN
Checking an agent demo: “Did it finish the task?”
Checking an agent workflow: “Did it update the right customer? Was the approval still valid? Did the payment fail, or just the confirmation? And why is it trying again?”
@16vchq We're building Praxis for teams automating business workflows.
Working through how to make human-AI alignment clear in the interface: when AI can act and when it needs approval.
Looking to connect with operators dealing with these decisions every day.
@ganesh2412@i_mika_el@Ingvar29007836@Michaelchoong5 The operator also needs to see whether they're viewing the original decision or a new evaluation. If today's rules give a different answer, show both with their reasons. A replay shouldn't quietly replace the record of what was authorized at the time.
@acecreamu Exactly. If someone already contacted the client while approval was pending, the original next step could now mean duplicate outreach. Before acting, the workflow should check what's changed and ask for fresh approval if the action no longer fits.
A client inquiry lands in your inbox. An AI agent researches the company and writes a brief.
Then someone still has to check the sources, approve the next step and make sure the right record gets updated.
That’s the workflow your team needs to automate.
Before choosing an agent tool, test what happens when:
→ Two companies have similar names.
→ Required information is missing.
→ Approval arrives a day late.
→ The destination app goes down.
→ The same inquiry triggers twice.
The automation needs checks that establish whether the work met your requirements, plus a clear path for anything unresolved.
Our guide compares Praxis, LangGraph, CrewAI, n8n and Langfuse through one practical example: turning an inquiry into an approved company brief and an assigned next action.
Each serves a different role, from custom agent execution to app connections and evaluation.
With Praxis, we’re building around the operator: the people, records, business rules and approvals needed to guide the work.
Read the guide: https://t.co/TK1muB1Lak
@Orbyntis And if the brief changes after outreach is queued, that approval needs another check before sending. The recruiter should see what's changed and which messages are affected, so candidates don't receive yesterday's offer under today's brief.
1/3
"Automating a recruiting agency is easy."
Collect CVs. Match candidates. Book interviews.
The reality:
> CVs arrive across three inboxes
> The client changes the brief
> Feedback sits in someone's DMs
> A candidate gets the same follow-up twice
@colbyshir3 A one-button approval still needs the verified supplier, destination and source invoice in view. Adding a supplier is a separate consequential step. What evidence would you want before approving the first payment?
@omarltx11@invflight For invoice validation, a useful exception is an amount that extracts cleanly but doesn't match the line items. What does the reviewer see when those two disagree?
@GlobalEllah The monthly discrepancy review needs an owner and a place to keep the supporting records. Otherwise the same unexplained items roll into next month. Which record do you find small-business owners struggle to keep current?
@HiceyIcecream Your comparison makes clearance capacity part of the buying decision. I'd give every missing document a named owner before shipment. Which paperwork issue most often delays a distributor's first direct order?
@FutureConsLtd Customers, payments and records each have a different finish line. I'd define one completed outcome and its exception owner before connecting tools. Which recurring task do your business clients ask you to fix first?
@abidsami_ Waiting between retries helps with availability; it doesn't prove the first request did nothing. Which actions must be checked before retrying?