Tandem's agents watch every store, find what's wrong, write the fix, send it to the right person and check the proof. You tap once. Store teams get a shorter, clearer day.
- 0
- Jobs agents finished
- 0
- Stores fixed
- 0
- Taps it took
Illustrative demo · each square is a store · tap approve yourself
Watch a normal day break. Then watch agents handle it.
Five moments every big chain knows. Keep scrolling: each one plays out the way it does today, then again with agents doing the work and a person tapping approve.
- 07:02. Without Tandem: One store's shelves have been emptying for three days. The numbers sit in a report nobody opened. The store is busy, HQ is busy, and sales leak until someone happens to look. Human effort: hours of digging, if anyone looks. With Tandem: An agent spotted it, wrote the fix and asked once. Signal Watch reads every store against its own normal each morning, drafts the task with the evidence and waits for a tap. The store gets clear steps, not a report. Human effort: one tap.
- 09:40. Without Tandem: The dashboard says 95% done. The floor disagrees. A ticked box is not a finished display. Forty proof photos sit in an inbox that nobody has time to open. Human effort: 40 photos to open, one by one. With Tandem: An agent checks every photo the moment it lands. Photo Verifier compares each photo with the task and sends misses back with the reason. Execution Chaser nudges the stores running late. Nobody at HQ lifts a finger. Human effort: none, it runs on its own.
- 11:15. Without Tandem: Cooler 2 starts warming. The paper log says it's fine. The last entry was Friday. Someone notices at close, a full afternoon of product later. Human effort: a ruined afternoon and a write-off. With Tandem: An agent caught it at the first bad reading. Compliance Agent opens the fix with the evidence, drafts the work order for the equipment and hands a six-minute product check to someone certified and on shift. Human effort: one tap.
- 14:00. Without Tandem: A new hire has a customer waiting and a question. The answer is somewhere in a binder in the back office. She guesses, or leaves the customer standing there to go and ask. Human effort: a walk to the back office and a lost sale. With Tandem: She asks out loud and has the answer in seconds. Ask Tandem answers from your own playbook, with the source, in her language. The store's own Store Brain remembers it, and anything it can't answer goes to HQ as a gap to fix. Human effort: one question.
- 18:30. Without Tandem: The holiday set is done. Did it sell anything? Completion rate is the only number HQ gets back. Someone builds a spreadsheet nobody trusts, and next quarter's plan is a guess. Human effort: a week in a spreadsheet. With Tandem: The result is measured for you, against control stores. Outcome Attribution compares stores that did the work with matched stores that didn't, inside the same district, so you see what the work earned. Human effort: none, it's already measured.
No inbox. No binder. Ask, snap, done.
The agents do the sorting, the chasing and the checking, so the store team gets one short list and a phone that answers back.
Illustrative demo data, including the sample policy answer.
Now picture that in every store.
HQ sees the whole fleet live: what the agents caught, what they already handled, and the few things waiting for a tap.
Illustrative animation: a fleet of 40 stores where agents sense a problem, decide on a fix, assign it, and prove it was done, such as a cooler reading 44F fixed and photo-verified in 18 minutes.
agents on one loop: sense, decide, do, prove
labor and safety rules checked against the official source
tokens to answer a question the store has already asked
MCP tools so your own assistant can work through Tandem
languages for associates, by voice or text
products on one data model and one audit log
Built for chains of 20 to 2,000 locations
- Specialty retail
- Grocery
- Convenience
- Quick service restaurants
- Multi-site services
Most store software sends work. Tandem closes the loop.
Comms apps stop when the message is read. Dashboards stop at the chart. Tandem runs Sense, Decide, Do and Prove every day, for every store.
Every store, read against its own normal.
Agents read POS, inventory, labor, punches, temperatures, fixtures, photos and what associates say, store by store. A drift is judged against that store's own Tuesday, not a fleet average.
The fix is drafted, with evidence and a price in minutes.
Each finding becomes a proposed action with the data behind it, a confidence score and a labor estimate. People approve it in the Agent Inbox, or the agent acts on safe work if you've allowed it.
Out-of-stock rate rose 98% above baseline at Store 207 over 3 days. Other stores are steady, so this is execution, not supply.
Work lands where it fits.
Tasks go into each store's daily plan, sized to the hours that store actually has, assigned to someone qualified and on shift, in the associate's language.
- Fix out-of-stocks45 min
- Holiday table, 3 levels60 min
- Cooler 2 product check6 min
- Price-change batch38 min
Proof, then outcome.
Photo Verifier checks the proof against the task's criteria. Attribution compares stores that did the work with control stores, so you see what it moved.
- Three height levels visible
- Table at the entrance
- Signage straight and readable
One system for the frontline.
31 products on one data model: the same stores, agents and audit log. Start with one; the rest are already connected.
- Cold holding logpassed
- Sanitation checklistpassed
- Fire extinguishersdue 4pm
- Traceability recordcomplete
- 09:41agentaction.create_task
- 09:43useraction.approve
- 10:02usertask.complete
- 10:02agentphoto.verify pass
Previews use illustrative demo data.
From HQ to the sales floor, one plan.
Send work stores can actually do, and see what it earned.
See the Command CenterPaste a brief and get store-ready tasks with labor minutes, photo criteria and targeting.
See the hours across the fleet and which stores a campaign pushes over budget, before it goes out.
Approve agent-drafted fixes with evidence and confidence, or let safe work run on its own.
Compare stores that executed with control stores, matched within district.
A model for every store, owned by the store.
Each store trains a small model from scratch on its own data: its week, what's normal on a Tuesday versus a Saturday, how its team executes, and every answer it has needed.
- Repeat questions are answered locally for zero LLM tokens. Only new ones reach a model, and then they're learned.
- Anomalies are judged against the store's own normal, so a quiet store and a busy flagship each get their own thresholds.
- No pooled data and no pretrained weights. The store can export it, erase it or opt in to anonymized benchmarks.
Hands busy? Just ask.
Associates ask by voice or text and get a short answer from your playbook, with its source, in their language. Questions the playbook can't answer show up for HQ as content gaps.
Bring the assistant you already use.
Tandem runs an MCP server. Any MCP-capable assistant can read store status, the playbook and the Agent Inbox, and act through Tandem with the same governance as Tandem's own agents. Every call carries its key's scopes and lands in the audit log.
curl -X POST https://tandem-eight-pi.vercel.app/api/mcp \ -H "Authorization: Bearer $TANDEM_API_KEY" \ -H "Content-Type: application/json" \ -d '{"jsonrpc":"2.0","id":1,"method":"tools/call", "params":{"name":"get_store_status", "arguments":{"storeCode":"S102"}}}'Agents earn autonomy. They don't start with it.
Drafts actions. A person approves each one in the Agent Inbox.
- Draft tasks, messages and offers with evidence
- Wait for approval in the Agent Inbox
- Act without a person approving
You set the level per agent, see each one's acceptance record, and can lower it at any time.
- Stores per actionmax 25
- Task minutes per store per daymax 60
- Confidence to act alone75% or more
- Quiet hours22:00 to 06:00
- Allowed actionsper agent
- Duplicate suppression5 days
Default values shown; each is set per agent.
One control pauses every agent at once. Work already in stores stays; nothing new is drafted or sent.
Illustrative. In the product this is HQ admin only and logged.- 09:41:02compliance-agent proposed create_task Check product in Cooler 2 store S102
- 09:41:02compliance-agent guardrails passed 1 store 6 min confidence 0.93
- 09:43:17store manager approved action executed task created
- 09:58:40associate task.complete photo attached
- 09:58:44photo-verifier verdict pass score 0.92
- 10:02:09district manager viewed evidence and proof photo
- Tenant isolation enforced on every database query
- Role-based permissions with store-scoped approvals and a read-only inspector role
- Single sign-on with OIDC and SAML, plus SCIM provisioning
- Phone numbers encrypted at the field level; personal data stripped before any model call
- Scoped API keys (read, write, approve, optionally limited by area), stored as hashes
- Replayable audit log of every human, agent and API action, and every agent action reversible
- Kill switch and per-agent guardrails
- Record retention by rule, with legal hold that overrides deletion
One platform instead of five.
Chains usually stitch together a comms app, a task app, a WFM suite, a compliance tool and merchandising analytics. Tandem runs them on one data model, so an agent can see a cooler reading, a schedule and a task budget at once.
| Capability | Store comms apps | Task apps | WFM suites | Compliance tools | Merch analytics | Tandem |
|---|---|---|---|---|---|---|
| Targeted messages with read and acknowledgement | partial | |||||
| Tasks with photo proof and AI verification | partial | partial | ||||
| Labor check before HQ sends work to a store | partial | |||||
| Schedules, time and leave | ||||||
| Fair workweek and overtime rules in code, with sources | partial | |||||
| Temperature logs, HACCP and inspection packets | partial | |||||
| Recall pulls sized in labor minutes | partial | |||||
| Sales mapped to fixtures | ||||||
| In-store experiments against control stores | partial | |||||
| Agents that act inside guardrails you set | partial | partial | ||||
| Per-store model that answers repeat questions for 0 tokens | ||||||
| MCP server for outside assistants | partial |
Categories describe typical products in each space, not any specific vendor.
It reads what your stores already produce.
- POS
- Inventory
- Labor hours
- Punches
- Temperatures
- Proof photos
- Planograms
- Fixture sales
- Recalls
- Certifications
- Traffic
- Associate voice
- Store walks
- Tickets
- Surveys
- Shrink
Start with the loop. Add the rest.
Every tier includes governance, the audit log, the API and the MCP server. Pricing depends on store count; talk to us.
Run the whole store on one system.
- Agents
- Execute
- Know
- Grow
- Staff
- Comply
- Operate
- Platform
Questions buyers ask.
What happens to our data, and is it pooled across stores?
Each store's Store Brain trains only on that store's data. It isn't pooled with other stores, and the store can export it as JSON, erase it, or opt in to anonymized benchmarks. Every train, export and erase is written to the audit log.
Does Tandem need an AI model to work?
No. Every agent has a deterministic engine, so the product runs with no model key. With a key, Tandem uses Anthropic's Claude for drafting, triage, translation and photo checks. Repeat questions are answered by the Store Brain without a model call.
How are agents governed?
Each agent has one of four autonomy levels: Observe, Propose, Act on safe work, or Autonomous. It also has guardrails on stores per action, task minutes, confidence and quiet hours. District managers approve only actions on their own stores, a kill switch pauses every agent at once, and every action is in the audit log.
Does AI decide whether something is legal or compliant?
No. Labor law and compliance checks are deterministic code. Each rule is stored with its official source and the date it was last verified, and unverified rules are shown but never enforced. Models only draft, extract and explain.
Can we keep our existing workforce management system?
Yes. Tandem has adapter interfaces for importing schedules from WFM suites such as UKG, Workday and Dayforce, and for exporting approved timecards to payroll. Connections are set up per customer during onboarding.
Which languages do associates get?
Ask Tandem answers in the associate's language. HQ messages can be translated for each associate when a model is connected. The associate experience is built for English, Spanish, French and Canadian French.
Does it work when a store loses signal?
Yes. Tasks, checks, temperature logs, photos and clock punches are queued on the device and sync when the connection returns. Work captured on a device keeps the device's own timestamp. Punches captured offline are flagged for manager review. Temperature logs and checks are append-only. If HQ already verified or rejected a task, the server's decision wins and the device's photo is kept as a note.
Can we try it before talking to sales?
Yes. The live demo runs on a fictional chain. Sign in as HQ, a district manager, a store manager or an associate and use the product as that role.
See Tandem run on your own stores.
Send a week of sales and your store list. We'll show which stores the agents would have flagged and what the work would have cost in labor.