Self-serve is in early access

Every answer shows its work.

Ward connects read-only to the systems you already run. Every figure in an answer links to the row it came from, and Ward checks it before you see it.

Read-only by default No data migration 5 TB on a 16 GB server Bring your own model
Finding · shrink-watch agent · 06:47

Store #37 showing 4.2% shrinkage vs 1.8% estate average. Pattern suggests receiving dock discrepancy, not shoplifting.

Proof chain
Source rowsinventory_adjustments, POS export (synced) · 9,412 rows
Queryshrink units ÷ received units · store 37 vs all stores · week 45
  • Valid data
  • Efficient query
  • Answers the question
Full scan rejected. Query rewritten, checks run again.
Sent to Store operations, awaiting approval
Store operationsTo review
4.2% · Shrinkage DetectionProof attached
Supply chainTo review
MerchandisingTo review
  1. 01Rows compress into the lake
  2. 02Agents scan, the index skips the rest
  3. 03A finding
  4. 04The proof chain checks it
  5. 05Sent to the team that acts

Sample grocery dataset. Team boards stand in for the project tool your team already uses.

The proof chain

Hover any number. See where it came from.

Every number a business acts on should be provable. Teams lose weeks arguing over whose figure is right, and that time belongs to deciding what to do. So a figure in a Ward answer is a citation, not text. It links to its source rows and its query, and it passed three checks before the answer came back.

Ward chat Sample grocery dataset
YouWhy is dairy margin down this week?
Ward

Dairy gross margin fell to this week, from the week before. Most of the drop is shrink at Store #37, which ran against a estate average. Receiving logged short at the dock across four deliveries, so the cause looks like receiving, not theft.

Hover or tap any number to open its proof chain.
21.4% Dairy gross margin, week 45
  1. 1Source rows
    POS export · synced into the lake · transactions ⋈ item_costs · 1,836,410 line items read
    storeskunet salescost
    3741022 Whole milk 1 gal$4.29$3.31
    5241870 Greek yogurt 32 oz$5.99$4.48
  2. 2Query
    from
    transactions, item_costs
    filter
    department = dairy · week 45
    group
    none
    measure
    (net sales − cost) ÷ net sales
  3. 3Checks
    • Valid data1.84M rows, no nulls in cost, week 45 complete
    • Efficient queryRead 1 of 52 week partitions
    • Answers the questionMargin for dairy, this week: matches the question
Returned only after all three passed.
24.0% Dairy gross margin, week 44
  1. 1Source rows
    POS export · synced into the lake · transactions ⋈ item_costs · 1,791,055 line items read
    storeskunet salescost
    3741022 Whole milk 1 gal$4.29$3.27
    1440115 Butter 1 lb$4.79$3.52
  2. 2Query
    from
    transactions, item_costs
    filter
    department = dairy · week 44
    group
    none
    measure
    (net sales − cost) ÷ net sales
  3. 3Checks
    • Valid data1.79M rows, week 44 complete
    • Efficient queryRead 1 of 52 week partitions
    • Answers the questionThe comparison week the answer names
Returned only after all three passed.
4.2% Shrink rate, Store #37, week 45
  1. 1Source rows
    POS export · synced into the lake · inventory_adjustments · 9,412 adjustment rows read
    storeskureasonunits
    3741022 Whole milk 1 galshrink−18
    3741870 Greek yogurt 32 ozshrink−11
  2. 2Query
    from
    inventory_adjustments
    filter
    store = 37 · department = dairy · week 45
    group
    store
    measure
    shrink units ÷ received units
  3. 3Checks
    • Valid dataAdjustment reasons all in the shrink code set
    • Efficient queryClustered on store, so one store read is one range
    • Answers the questionStore-level shrink, the figure the answer cites
Returned only after all three passed.
1.8% Shrink rate, estate average, week 45
  1. 1Source rows
    POS export · synced into the lake · inventory_adjustments · 402,977 adjustment rows read
    storeskureasonunits
    1440115 Butter 1 lbshrink−4
    5241022 Whole milk 1 galshrink−6
  2. 2Query
    from
    inventory_adjustments
    filter
    department = dairy · week 45
    group
    none
    measure
    shrink units ÷ received units
  3. 3Checks
    • Valid dataEvery store reported for week 45
    • Efficient queryRead 1 of 52 week partitions
    • Answers the questionSame measure as Store #37, so the two compare
Returned only after all three passed.
312 cases Cases short at receiving, Store #37, week 45
  1. 1Source rows
    SAP · read-only replica, queried in place · goods_receipts ⋈ purchase_orders · 1,204 receipt lines read
    POdeliveryorderedreceived
    918274Tue 06:1014096
    918390Thu 05:5512081
  2. 2Query
    from
    goods_receipts, purchase_orders
    filter
    plant = 37 · department = dairy · week 45
    group
    delivery
    measure
    ordered − received, in cases
  3. 3Checks
    Attempt 1 rejected on efficient query. It read every plant's receipts for the year. Ward rewrote it with plant and week filters and ran the checks again.
    • Valid dataEvery receipt joins to a PO
    • Efficient queryRewritten with plant and week filters
    • Answers the questionShortfall at the dock, which is what the answer claims
Returned only after all three passed.
Check 01

Valid data

The rows exist, the types match the measure, and the window is the one asked about.

Check 02

Efficient query

The query uses the partitions and indexes it should. A full scan where a filter would do is rejected.

Check 03

Answers the question

The result is checked against what was asked. A correct number for a different question does not pass.

On a fail

It loops back

A failed check is not shown to you as a warning. The query is rewritten and checked again, and only a figure that passes all three is returned.

When the data is thin

“What did the Easter promo do at Store #52?”

I can't answer that from what is connected. Store #52 has no promo calendar rows before March, so there is no Easter window to measure. Try one of these:

Promo lift at Store #52 since MarchEaster promo lift across all storesWhich stores are missing promo calendar rows

No guess. A dead end comes back as the next question.

How the proof chain works →
Query in place

Ask across every system you run. Move nothing to do it.

Live databases connect through Ward’s proxy and are joined where they sit, so a read-only replica is all Ward needs. Sources that are not databases, like exports, Stripe, logs and PDFs, sync into the lake. Query in place or sync: your choice, per source.

Live databases
Postgreslive databaseno data moved
SAPread-only replicano data moved
Files and SaaS
POS exportfilessynced
StripeAPIsynced
Ward
ProxyJoins live tables where they sit. A read-only replica is enough.
Data lakeColumnar, indexed, partitioned and clustered.
One queryAcross every source, as one database. Each figure linked to its row.

FreshnessSynced sources land by batch, micro-batch or streaming. Streaming rows are queryable the moment they arrive.

The engine

Compressed for the compute bill. Built so the model never does the math.

Most AI analysts hand a model a pile of rows and hope the arithmetic holds. Ward’s own data lake runs every calculation and hands the model the result. Smaller context means a smaller model bill, and less for the model to make up.

5 TBof data in one live deployment
5Brecords, growing toward 100B
16 GBserver it is queried on
0migrations to get there

Columnar and compressed

Parquet-based storage on a custom fork of Apache DataFusion. Columns compress far better than rows, so the same data takes less disk and less memory to scan.

Its own index format

Ward keeps its own indexes for lookups and query planning. A question about one store and one week reads that store and that week, not the table.

Partitioned and clustered at once

The same table is partitioned and clustered together, so a query can skip whole partitions and then read a narrow range inside the one it keeps.

The model never does the math

The AI turns your question into queries and writes the answer. The lake computes every figure. The model sees the question, the schema and the result, never your tables, so there is less to pay for and less to make up.

The figures above are one live enterprise grocery deployment, not a benchmark. The full architecture, for the people who will review it →

Access control for data and agents

An agent sees what the person who runs it can see. Nothing more.

Access is written as row- and column-level policy and applied on every query, below the model, so no prompt can widen it. People and agents answer to the same rules, and every read is logged with the reason it was allowed or denied.

01
Row and column policies. Access is written as Cedar policies at row and column level and applied in real time, on every query, for every caller.
02
Agents inherit the person. An agent runs with the permissions of the user who starts it. It cannot see a row that person could not see, whatever the prompt says.
03
Every access, with its reason. Each read is logged with the policy that allowed or denied it. A reviewer reads why, not just what.
04
Isolated per tenant. Customer data is physically separated per tenant. Only schemas and settings are gated in the application.
05
Models under policy. You choose the model providers. A policy decides who is allowed to change them.
06
Where it runs. Your cloud, ours, or on-prem, including air-gapped with scheduled updates. SSO through OpenID Connect.
How access control works →
Access logsample
06:47:12 shrink-watch for m.kim inventory_adjustments allow region = North, as m.kim
06:47:12 shrink-watch for m.kim payroll.hourly_rate deny column outside store-ops policy
06:48:03 availability for j.osei stock_levels (SAP) allow supply-chain role, all stores
06:52:40 r.patel models.provider deny only platform admins change models

One log for people and agents. The second line is the same agent, the same person, one column further, and the policy that stopped it.

Not a mockup

One take, from an empty source list to a model in the graph.

Recorded against a running tenant. Nothing is sped up and nothing is cut except the login screen.

Ward · connecting a POS export 63s · one take
Describe the export in a sentence. Ward drafts the source, you apply it, the pipeline runs, and the model joins the graph. 63 seconds, unedited, one take. Only the login screen is cut.
Where it is running

One enterprise deployment, and the arithmetic behind the rest.

What is running and what is modelled sit in separate boxes, so you can judge each on its own.

Enterprise grocery · nine-figure revenue

Live against POS, ERP and inventory.

5B records, growing toward 100B. 5 TB, queried on a 16 GB server.

Ingestion, the data lake and the reasoning on top, running on production systems. Stockouts, shrink, fill rate and assortment are where it starts, because that is where the dollars are.

We do not name operators or publish their numbers without permission. Reference calls get arranged while you are still evaluating. How the engagement is going →

Modelled arithmetic

Where 200 basis points would come from.

Four signals, worked out for a mid-market multi-store estate. Nobody has booked a dollar of it yet.

Fill rate, fewer lost baskets~30 bps
Assortment, better mix and less long tail~50 bps
Shrink, cause-attributed loss~80 bps
Promo, cannibalisation caught mid-flight~40 bps
Modelled EBITDA opportunity~200 bps

Retail is where Ward is deployed, not what Ward is. The retail build →

Compare

Already evaluating something else?

Each page gives the other product its credit first, then answers the same question both ways with the proof chain visible.

Pricing

Free to deploy. The number is on the page.

Platform
Free
The whole console, deployed. No license, no seats, no platform fee.
  • Every connector and the semantic layer
  • Unlimited agents and users
  • Row and column policies, access log, SSO
  • Proof chain checks on every answer
Request access
Setup
From $0
Self-guided is free. Paid packages when you want a named engineer for training and guidance.
  • Self-guided: $0
  • Guided, two weeks: $7,500
  • Managed, six weeks: $20,000
  • Program, multi-entity: from $50,000
See the setup packages

Verification is included. Every figure is checked on a small context window, so the checking adds little to the model bill. No seats, no license, no annual plan. How the 5% is metered, and the setup packages with a named engineer, are on the pricing page.

Questions

What a technical evaluation asks first.

Proof and accuracy

The model never does the math. It turns your question into queries and writes the answer, and Ward's data lake computes every figure. Each figure in the answer is a citation linked to its query and source rows, and it is checked three ways before the answer returns: the data is valid, the query is efficient, and the result answers the question asked. A figure that fails is not shown with a warning. The query is rewritten and checked again. When the data is too thin to answer, Ward says so and suggests questions it can answer.

Very little. Verification is included, and because the lake computes the figures, the model works on small context windows: the question, the schema it needs and the query results, never your tables. That keeps both the answer and the checking cheap. You pay 5% of the model compute that routes through Ward, on your own provider keys, and nothing per check.

You choose the model providers, and a policy decides who is allowed to change them. What a model sees is limited by the same row- and column-level policies as the person asking: it receives query results that person is allowed to see, never whole tables. Routing specific tasks to specific models is on the roadmap.

Cost and footprint

Nothing to deploy. The console, every connector, every agent and every user are free. Ward bills 5% of the model compute that routes through it, metered on your own provider keys, and that is the only recurring fee. Setup is a one-time package if you want a named engineer for training and guidance, and the smallest one is $0. There is no license and nothing is priced per seat.

No. Live databases connect through Ward's proxy and are joined in place, so a read-only replica is all Ward needs and nothing is moved. Sources that are not databases, like exports, Stripe, logs and PDFs, sync into Ward's data lake. Query in place or sync: your choice, per source. Revoking the credential returns you to exactly where you started.

Depends what you run today. Ward includes ingestion, its own data lake engine, transformation, the semantic layer, query and chat, and agents under policy, so the first things teams retire are usually a managed pipeline tool and a separate metrics layer. Columnar storage with its own indexes keeps compute small: one live deployment queries 5 TB and 5B records on a 16 GB server. It does not replace your source systems, your CDP or your data science platform.

Agents and security

What the person who runs it can reach, and no more. Agents run with the permissions of the user who starts them, under the same row- and column-level Cedar policies, applied in real time below the model, so a crafted instruction cannot widen them. Every access is logged with the policy that allowed or denied it.

Only with an explicit grant, per table and per action, and you can require a human approval on each one. Findings are routed to the team that acts on them, awaiting approval, rather than acted on silently. Read-only is the default and most teams never leave it.

Where you choose: your cloud, ours, or on-prem, including air-gapped deployments with scheduled updates. In Ward's cloud, customer data is physically isolated per tenant; only schemas and settings are separated in the application. Sign-in is SSO through your OpenID Connect provider.

Connect one source and see what comes back.

Read-only, and nothing migrates. Hover the first number Ward gives you and check it yourself. Deploying is free; you pay 5% of the model compute it uses.

Find out what your data has been hiding.

Tell us about your operation. We’ll show you the problems Ward catches, and the ones your current tools miss.

Step 1 of 3
What are your goals?
Step 2 of 3
About your operation
Step 3 of 3
Your contact info