Pilot proposal · SmartBiz · September 2026
For the Smart Commerce team
Hawky Agents ×
Amazon Smart Commerce
A Creative Agent and a Performance Agent that do an agency's job for D2C brands on SmartBiz — switched on from the console, callable from Amazon's own seller agent.
Contents
01 · Who is Hawky
The agentic performance marketing platform
Always-on agents that read what performs, produce on-brand creative from the evidence, and keep iterating against the KPI. Humans own strategy and brand. Hawky owns the production and the iteration.
Brands & agencies
Ads analysed
Creatives / month analysed by Hawky
Creatives / month, single brand
Creative performance, proven
Trusted by 100+ brands and agencies
“It's not just the quantity I love, it's the quality. I uploaded my brand book once and every ad speaks the language we want to speak.”
Pooja Jena · Brand Marketer · Cars24 — 1,000+ creatives / month · 15–20% lower CPA02 · The problem
Paid is the front door.
For a new D2C brand, performance marketing is not one channel among several. It is most of the traffic, and the seller who cannot run it well does not get found.
Of a new brand's traffic comes from paid
Of ad spend to an agency, negotiated
The same fee, un-negotiated
Paid is the traffic
A brand with no audience of its own buys every early visitor. Stop the spend and the storefront goes quiet.
The fee grows with the budget
An agency is paid a share of spend, so each rupee added to the budget adds 7 to 20 paise of fee — whether or not that rupee worked.
The good ones say no
Finding a good agency at an SMB budget is hard. The seller at ₹1–2 Cr monthly TPV pays the premium or settles for whoever says yes.
Agency rates from the September call. Hawky line illustrative — commercials follow the pilot read-out.
The sellers who most need performance marketing are the ones least able to buy it well. That is the gap SmartBiz can close, and the one the rest of this deck is about.
03 · Agents, not an agency
Two agents do the agency's job.
The Creative Agent and the Performance Agent split the work an agency team does, and run it every day against the seller's own data instead of once a week against a report.
The loop an agency runs weekly, run every day against the seller's own orders.
| An agency | Hawky's agents | |
|---|---|---|
| How it is paid | 7–10% of ad spend, up to 20% | Slabbed plans. Not a cut of spend. |
| New creative | Days, through a brief | Same session, from the catalogue |
| Creative volume | What the retainer covers | Per SKU, per occasion, per region |
| Account attention | Weekly call, business hours | Read every day |
| Optimises to | Pixel events | Delivered, non-returned orders |
| How it scales | More brands means more people | Easily scalable agents, slabbed pricing |
Automate
The routine runs itself
Production, launch, reallocation and reporting run without a person in the loop. The seller only approves.
Streamline
One switch, one question
No brief, no agency onboarding, no weekly call. Margin is the only thing the seller is ever asked.
Cheaper
Software, not a cut
Priced as a plan rather than a percentage, so growing the budget does not grow the fee.
Reliable
Built not to guess
Numbers computed in code, fixed e-commerce sequences, tenant isolation, and nothing spends without approval.
04 · What Amazon needs
The engine is the open question.
Four things came out of the September call. This deck answers them in order, and nothing else.
01 · Cost
Below an agency
Marketing a ₹1–2 Cr TPV seller can afford, priced so the fee does not rise every time the budget does.
02 · Scope
Engine, not interface
Seller experience is Amazon's to build. What is open is whether the engine behind it is built or partnered. This proposal is for the engine and the connector into it.
03 · Effort
Nothing to configure
Sellers will not fill in a brief, set a bid strategy or name an audience. One switch in the console, and the work starts.
04 · Alpha
Beyond catalog ads
Catalog already carries 60–70% of spend. The gain has to come from trigger and seasonality creative and from whitespace, not from every seller chasing the same competitor signal.
Amazon owns the seller experience. Hawky owns the marketing work underneath it. Everything below describes what sits under the switch, and the two ways it connects.
05 · What sits under the switch
Context first, then the decision.
The agents do not reason over raw accounts. Query pipelines aggregate the data deterministically; the model only decides and explains.
Brand context
Ad accounts, product-level performance, attribution and bottom-funnel data. Inside SmartBiz this also means catalogue, theme tokens, pricing and delivered orders.
Competitor context
Category creative patterns, scaling behaviour, spend signals and test cadence — read for the gap in the category, not for the thing everyone is already buying.
Trend context
Seasonality, occasions and cultural moments, mapped to the SKUs in the catalogue that can actually carry them.
Arithmetic is never left to the model. Agents write code on the fly to structure and serialise the data, the numbers are computed in the pipeline, and the model receives a table. Sequences are hardcoded for e-commerce, and each brand's context is isolated — agents are blind to every other tenant.
Context in, computed table out. The model is the last step, not the first.
06 · Creative production
The catalogue is the brief.
SmartBiz already holds the product images, the copy, the pricing and the brand's theme tokens. That is everything a creative brief normally asks a seller to write.
Input
Zero-brief start
Catalogue plus theme tokens produce the first set of on-brand statics without the seller writing anything.
Variation
Occasion and region
The same SKU rendered for a festival, a city, a weather window or a price change — the variants a catalog feed cannot express.
Prospecting
Trigger-based
Lifestyle and occasion-led assets for the top of funnel, where native catalog ads have the least to say.
Retirement
Fatigue before the drop
Hook and hold rate flag a tiring asset days before CPA moves, and a replacement is queued.
Cars24 is the closest existing analogue. 160–180 vehicles uploaded a day, region-specific creative generated automatically against a moving catalogue, now running in Dubai and Australia as well as India.
07 · Performance operations
The agent decides what. The seller decides when.
Nothing runs proactively on day one. The agent proposes, the seller approves in one tap, and the scope of what it may do on its own widens as the results land.
Recommend
Every action arrives as a proposal with the numbers behind it. The seller approves or ignores. Nothing spends without a tap.
Approve once
Classes of action the seller has approved repeatedly — pausing a dead ad set, shifting budget between two live ones — move to standing permission.
Run and report
The agent operates inside a budget ceiling and reports after the fact. The seller keeps the ceiling and the off switch.
Inside SmartBiz the agent can optimise toward a delivered, non-returned order. Order status, returns and RTO sit in the same system as the ad data. An external agency optimises to a pixel event and never sees what came back.
08 · Brand onboarding flow
Brand onboarding flow.
From switch-on to first spend in one session. Four lanes, five phases, and the brand acts three times: switch on, one number, one tap.
Read left to right. Orange boxes are the only moments the brand is asked for anything.
The brand does
Three things
Switches on and authorises Meta and Google, enters gross margin if the data cannot show it, and approves the first plan.
SmartBiz passes
Store data
Catalogue, theme tokens, pricing and order history — then delivered and returned orders, every day after.
Hawky builds
Context, then a plan
Brand context and a category read in FeatherDB, targets from margin, five creatives and a campaign structure.
Then it runs
A daily loop
Delivery and order outcomes are read every day. Spend is reallocated, fatigued creative is replaced, and a report lands monthly.
A brand that has never run an ad should be live the same day, having made one decision. Everything in the Hawky lane exists so the brand lane stays this empty.
09 · Where it runs
One adapter layer, every channel behind it.
The agents never talk to a platform directly. Each channel sits behind an adapter, so onboarding a new platform means writing one adapter, not rebuilding the agents.
The agents never call a platform directly. Everything goes through an adapter.
Live
Meta and Google
Full read and write: structure, budgets, bids and creative, managed by the Performance Agent.
Other regions · in progress
TikTok and LinkedIn
TikTok runs for brands outside India. LinkedIn is in progress on the same adapter pattern.
Hybrid
Purplle, Myntra, Zepto
Activity logs and reports are uploaded and read into the same context. Recommendations come back out.
Agent to agent
The Hawky Connector
The agents exposed over MCP, so another agent — Seller Assistant or Amazon's own engine — can discover and call them.
Amazon can point its own engine at the Hawky Connector today. It is the fastest way for the SmartBiz team to see what the agents can do, before a single seller is onboarded.
10 · How it connects
The product and the Connector, in parallel.
Not a sequence. Both tracks start in the pilot, and both read the same context contract, so a seller gets the same answer in either surface.
Track 1 · The product
Hawky's creative and performance modules rendered inside the SmartBiz console, white-labelled. Amazon keeps the surface and the seller relationship; Hawky runs the work. This is what the pilot sellers use on day one.
Track 2 · The Hawky Connector
The same agents exposed as tools through the Hawky Connector, built on the open MCP standard. Amazon's own seller agent connects, discovers the capabilities and calls them. Nothing is rebuilt for it — it is the same engine behind a different door.
One contract
Catalogue, orders, ad accounts and margin in; creative, campaigns and recommendations out. Both tracks read and write the same shape, so the two never drift apart.
The product proves the outcome. The Connector proves the integration. Running them together means the pilot answers the commercial question and the architecture question in the same month.
Both tracks start in the pilot. Neither is a rebuild of the other.
11 · Amazon's seller AI
Seller Assistant asks. Hawky answers.
Amazon India's Seller Assistant is already agentic, already free to every seller, and already able to act with permission. It does not need a marketing tab. It needs a marketing skill it can call.
Read
get_account_read
Returns the current state of spend, delivery and creative performance for a seller, computed in the pipeline, not narrated by a model.
Make
generate_creative
Produces on-brand statics for a SKU, an occasion or a region from the seller's own catalogue and theme.
Launch
launch_campaign
Builds structure, budget and bidding on Meta and Google and puts the campaign live, subject to the seller's permission.
Decide
recommend_actions
Returns ranked actions with the evidence attached, so the assistant can present them in its own voice and its own approval flow.
The seller keeps talking to Amazon. Seller Assistant stays the surface, keeps the permission model and keeps the conversation. Hawky is the capability it discovers when a seller asks a question about growth that today has no answer behind it.
The seller never leaves Amazon. Permission and voice stay with Seller Assistant.
12 · The concerns raised
Answered before the pilot.
Four questions came up on the call that a pilot should be designed to settle, not to postpone.
Raised
Whitespace, not the herd
Competitor intelligence that everyone acts on inflates CPMs. The competitive read is built to return the gap in the category — the angle, occasion or format nobody in the set is running.
Raised
Alpha over catalog ads
We do not claim to beat a catalog feed at retargeting. The uplift is in seasonality adaptation and trigger-based prospecting, which is where catalog creative has the least to say.
Raised
RTO is the objective
Because SmartBiz holds delivered and returned order status, RTO becomes an optimisation target rather than a post-hoc report. Cohorts that return are learned against, not just counted.
Raised
Hallucination
Aggregation and arithmetic happen in code. The model sees a computed table and produces a decision and a summary. Agent sequences are fixed for e-commerce, and tenants are isolated.
13 · Proof
Catalogue creative, already at scale.
The closest thing to a SmartBiz seller in Hawky's book is Cars24 — a moving catalogue, a daily refresh, and creative that has to be produced without anyone briefing it.
Listings entering creative production daily
Markets live — India, Dubai, Australia
Briefs written per creative
Case · Cars24
Region-specific creative against a catalogue that changes every day.
Catalogue refresh cycle
Variant generation by region
Markets from one pipeline
A SmartBiz seller is the same problem at a smaller size. A catalogue that moves, no one to write a brief, and spend that has to work from the first week.
14 · The pilot
Three to four brands. One month.
Each brand is onboarded in a day, with its intelligence built and both agents ready to perform by the end of it. The rest of the month is the agents working, and Amazon's performance team checking them.
Brand onboarded, intelligence built, agents ready
Brands across maturity buckets
To a per-brand read-out
Day one carries the setup. Everything after it is the agents at work. Columns are phases, not to scale.
Brand A · New
Has never run paid. Tests the one-day start from nothing. Success is spend live on day one and a clean first month, not ROAS.
Brand B · Running
Already buying media, most of it catalog. Tests the alpha question directly — trigger and seasonality creative alongside what they run today.
Brand C · Established
Has a team and a history. Tests whether the recommendations survive review by people who know the account.
Onboard
The brand switches on and authorises its accounts. Brand, competitor and trend intelligence is built the same day, the trailing 30 days of account history becomes the baseline, and both agents are ready with a first plan for approval. Connector access for Amazon's engine goes live alongside.
In market, reviewed
Agent-produced creative runs beside each brand's existing set. Amazon's performance team reviews every recommendation before it goes live.
Agents run
Daily reads, weekly reallocation, fatigued creative replaced. Brands that are comfortable move routine budget moves to standing permission inside a ceiling. Seller Assistant calls the Hawky Connector on one pilot brand.
Read-out
One document per brand against its baseline, plus an integration note on what it would take to put this in front of every SmartBiz seller.
Measure
Time to ready
Hours from switch-on to agents ready and a first plan in front of the brand. The target is one day.
Measure
Creative throughput
Assets produced per brand per week, and time from catalogue change to creative live.
Measure
Recommendation quality
Share of agent recommendations the reviewing team accepts, and the outcome of those executed.
Measure
Performance vs baseline
CPA, ROAS and CTR for the month against each brand's trailing 30 days.
No commercial terms are proposed in this document. The pilot is scoped to answer whether the engine works for SmartBiz sellers. Pricing follows the read-out, once there is a result to price against.
15 · The ask
Name the brands, open the sandbox.
Four things from Amazon, five from Hawky, and the pilot starts inside a fortnight.
From Amazon
- Three to four sellers — picked across the maturity buckets (new, running, established), with consent to run the pilot and share baselines.
- Sandbox access — a SmartBiz test account with catalogue and order data, so the context contract is built against something real.
- One engineer — a counterpart on the Seller Assistant side for the Connector handshake. A few hours a week is enough.
- A reviewer — Amazon's performance team validating recommendations before they go live.
From Hawky
- The engagement team — assigned from week zero, one point of contact.
- Pilot creative — produced and funded by Hawky for every pilot brand.
- Connector access — live for Amazon's own engine from day one.
- Day-one onboarding — each brand live, with intelligence built and agents ready, within a day.
- The read-out — one document per brand at day 30, plus an integration note.
Confirm the pilot brands and the sandbox in the first week of October, and the read-out lands thirty days after the brands go live.