$\lvert A\rangle \langle B \rvert$ is the tensor of a ket and a bra (well, duh). This means it is an element of the tensor product of a Hilbert space $H_1$ (that's where the kets live) and of a dual of a Hilbert space $H_2^\ast$, which is where the bras live. Although for Hilbert spaces their duals are isomorphic to the original space, this distinction should be kept in mind. So we can "feed" a ket $\lvert \psi\rangle$ from $H_2$ to the bra in $\lvert \phi\rangle\otimes \langle\chi\rvert \in H_1\otimes H_2^\ast$, and are left with a state in $H_1$ given by $\langle \chi \vert \psi\rangle \lvert \phi\rangle$. The usual use case for such a tensor product is when $H_1=H_2$ to construct a map from $H_1$ to itself, e.g. the projector onto a state $\lvert \psi \rangle$ is given by $\lvert\psi\rangle \langle \psi \rvert$.
In general, a tensor in $H_2 \otimes H_1^\ast$ corresponds to a linear operator $H_1\to H_2$. In the finite-dimensional case, these are all linear operators, in the infinte-dimensional case, this is no longer true, e.g. $H^\ast \otimes H$ are precisely the Hilbert-Schmidt operators on $H$.
In constract, a tensor $\lvert A\rangle\otimes \lvert B\rangle$ (also just written $\lvert A \rangle \lvert B\rangle$) in $H_1\otimes H_2$, although it corresponds to a bilinear map $H_1\times H_2\to\mathbb{C}$ by definition, is usually not meant to denote an operator, but a state. Given two quantum systems $H_1$ and $H_2$, $H_1\otimes H_2$ is the space of the states of the combined system (as for why, see this question).
The two different Hamiltonian forms represent different physical things. The first, given by
$$
\hat{H}_1\otimes\hat{I}_2+\hat{I}_1\otimes\hat{H}_2\tag{1},
$$
represents the free Hamiltonians of each system separately, and the second, given by
$$
\hat{H}_1\otimes\hat{H}_2\tag{2},
$$
represents an interaction term between two quantum systems. (So, e.g. the first might be the kinetic and potential energies of two individual oscillators, and then the second might represent the interaction Hamiltonian if, for instance, the two oscillators are connected by a spring).
Importantly, Hamiltonians of the form (1) evolve unentangled states into unentangled ones, but Hamiltonians of the form (2) can evolve unentangled states into entangled ones. To see this, consider the following calculations.
Consider two quantum systems described by Hamiltonians $H_1$ and $H_2$, whose eigensystems are separately described by
$$
\hat{H}_1|\psi_n\rangle = \epsilon_n|\psi_n\rangle,~~~~~\hat{H}_2|\phi_n\rangle = \mu_n|\phi\rangle.
$$
Let's consider case 1 first, in which the combined Hamiltonian of the combined system happens to be
\begin{equation}
\hat{H}_1\otimes\hat{I}_2+\hat{I}_1\otimes\hat{H}_2\tag{1}
\end{equation}
Then, the eigenstates of this operator are the the product of the eigenstates of the individual Hamiltonians, and the eigenvalues are sums of the individual eigenvalues, which we can show by computing
$$
\left(\hat{H}_1\otimes\hat{I}_2+\hat{I}_1\otimes\hat{H}_2\right)\left(|\psi_n\rangle\otimes|\phi_m\rangle\right)
=(\epsilon_n+\mu_m)\left(|\psi_n\rangle\otimes|\phi_m\rangle\right).
$$
(The details involve just using linearity.)
Now, given an arbitrary unentangled initial state $|\Psi(0)\rangle$, it can be written as the product of vectors expanded in the individual energy eigenbases as
$$
|\Psi(0)\rangle
=
\left(\sum_{n}a_{n}|\psi_n\rangle\right)\otimes\left(
\sum_{m}b_n|\phi_m\rangle\right)
=
\sum_{nm}a_nb_m
\left(|\psi_n\rangle\otimes|\phi_m\rangle\right).
$$
We can then get the full time-dependence by attaching the exponential factors to the eigenvectors in the usual way, yielding
$$
|\Psi(t)\rangle
=
\sum_{nm}e^{-i(\epsilon_n+\mu_m)t/\hbar}a_nb_m
\left(|\psi_n\rangle\otimes|\phi_m\rangle\right).
$$
Crucially, this state factors as
$$
|\Psi(t)\rangle=\left(\sum_{nm}
e^{-i\epsilon_nt/\hbar}
a_n|\psi_n\rangle\right)\otimes\left(
\sum_{m}
e^{-i\mu_mt/\hbar}
b_m|\phi_m\rangle\right),
$$
and so if the system started unentangled, it remains unentangled. This is purely a consequence of the fact that the Hamiltonian is a sum of single-system Hamiltonians because this is what leads to the eigenenergies being sums of the individual ones, which allows us to factor the exponential.
Now, to see that that doesn't work in the case of a Hamiltonian of the second form, given by
$$
\hat{H}_1\otimes\hat{H}_2\tag{2},
$$
we first note that that the product of eigenvectors is still an eigenvector, but the eigenvalues are now products of the individual eigenvalues, i.e.,
$$
\left(\hat{H}_1\otimes\hat{H}_2\right)\left(|\psi_n\rangle\otimes|\phi_m\rangle\right)
=(\epsilon_n\mu_m)\left(|\psi_n\rangle\otimes|\phi_m\rangle\right).
$$
If we again start with the initially unentangled state shown above, then the full time-dependent state is given by
$$
|\Psi(t)\rangle
=
\sum_{nm}e^{-i(\epsilon_n\mu_m)t/\hbar}a_nb_m
\left(|\psi_n\rangle\otimes|\phi_m\rangle\right),
$$
which can no longer be factored in general!
This can also be seen by exponentiating the Hamiltonians directly to get the unitary time-evolution operators. For Hamiltonians of the form (1), the time-evolution operator is
$$
U_1(t) = \exp\left(
-\frac{it}{\hbar}
(\hat{H}_1\otimes\hat{I}_2+\hat{I}_1\otimes\hat{H}_2)
\right)
=
\exp\left(
-\frac{it}{\hbar}
\hat{H}_1\otimes\hat{I}_2
\right)
\exp\left(
-\frac{it}{\hbar}
\hat{I}_1\otimes\hat{H}_2
\right),
$$
which is allowed because the two operators commute with each other. Furthermore, it is relatively straight-forward to show that this can be written as
$$
U_1(t) =
\left(
\exp\left(
-\frac{it}{\hbar}
\hat{H}_1
\right)
\otimes\hat{I}_2\right)
\left(\hat{I}_1\otimes \exp\left(
-\frac{it}{\hbar}
\hat{H}_2
\right)\right)
=
\exp\left(
-\frac{it}{\hbar}
\hat{H}_1
\right)
\otimes
\exp\left(
-\frac{it}{\hbar}
\hat{H}_2
\right).
$$
Thus, this time-evolution operator can then be seen to "conserve unentangled-ness". The other one does not factor in that same way.
Best Answer
If the particles never did interact in the past and won't ever interact in the future, then it indeed makes not much sense to model the system as a tensor product. Keep in mind that it does not matter how long ago the interaction happened as long as the particles are not otherwise disturbed - that's one of the points of the EPR 'paradoxon'.
However, care must be taken when dealing with identical particles: They are excitations of the same underlying field and should always be considered as having interacted.
In fact, the state space of identical particles is not the tensor product, but its symmetrization (in case of bosons) or anti-symmetrization (in case of fermions), ie only entangled states are considered.
These are just my thoughts and by no means authorative - anyone who deals with such issues on a more hands-on basis probably has a better grasp of the underlying issuess...Considering the unexplained downvote, a more technical answer to the problem:
In non-relativistic Quantum mechanics, the evolution of a system in governed by inital state and Schrödinger's equation.
If we start out in an un-entagled state and the Hamiltonian does not contain interactions between the different sub-systems, we will stay in an unentangled state and we can consider the sytems independantly.
More formally, let our inital state be
$$\psi = \psi_1\otimes...\otimes\psi_n$$
and the Hamiltonian given by
$$H = \tilde H_1 + ... + \tilde H_n$$ $$= H_1\otimes1\otimes...\otimes1 + 1\otimes H_2\otimes1\otimes....\otimes1 + ... + 1\otimes...\otimes1\otimes H_n$$
Then the state after time $\Delta t$ is given by
$$ \mathrm{e}^{-iH\Delta t}\psi = \mathrm{e}^{-i\tilde H_1\Delta t}\cdot...\cdot\mathrm{e}^{-i\tilde H_n\Delta t}\psi_1\otimes...\otimes\psi_n = (\mathrm{e}^{-iH_1\Delta t}\psi_1)\otimes...\otimes(\mathrm{e}^{-iH_n\Delta t}\psi_n) $$
where the first equality follows from $[\tilde H_i,\tilde H_j] = 0$.